When Does Coordination Require Centralization?

Ricardo Alonso Wouter Dessein

Northwestern University University of Chicago

Niko Matouschek
Northwestern University

First version: September 2005

This version: July 2006

Abstract
This paper compares centralized and decentralized coordination when managers are privately
informed and communicate strategically. We consider a multi-divisional organization in which
decisions must be responsive to local conditions but also coordinated with each other. Infor-
mation about local conditions is dispersed and held by self-interested division managers who
communicate via cheap talk. The only available formal mechanism is the allocation of decision
rights. We show that a higher need for coordination improves horizontal communication but
worsens vertical communication. As a result, no matter how important coordination is, decen-
tralization dominates centralization if the division managers are not too biased towards their
own divisions and the divisions are not too different from each other (e.g. in terms of division
size).

Keywords: coordination, decision rights, cheap talk, incomplete contracts

JEL classifications: D23, D83, L23

∗
We thank Jim Dana, Mathias Dewatripont, Luis Garicano, Rob Gertner, Bob Gibbons, Andrea Prat, Canice
Prendergast, Luis Rayo, Scott Schaefer and Jeroen Swinkels for their comments and suggestions. We would also like to
thank seminar participants at Bristol University, Chicago GSB, Columbia University, Duke University, Ecares, Essex
University, HKUST, Indiana University, Northwestern Kellogg, Oxford University, the University of British Columbia,
the University of Toulouse and conference participants at the 2005 MIT Sloan Summer Workshop in Organizational
Economics and the 2006 Utah Winter Business Economics Conference for their comments. A previous version of this
paper has been circulated under the title "A Theory of Headquarters."
1 Introduction

Multi-divisional organizations exist primarily to coordinate the activities of their divisions. To do

so efficiently, they must resolve a trade-off between coordination and adaptation: the more closely
activities are synchronized across divisions, the less they can be adapted to the local conditions
of each division. To the extent that division managers are best informed about their divisions’
local conditions, efficient coordination can be achieved only if these managers communicate with the
decision makers. A central question in organizational economics is whether efficient communication
and coordination are more easily achieved in centralized or in decentralized organizations. In other
words, are organizations more efficient when division managers communicate horizontally and then
make their respective decisions in a decentralized manner or when they communicate vertically
with an independent headquarters which then issues its orders?
This question has long been debated among practitioners and academics alike. Chandler (1977),
for instance, argues that coordination requires centralization:

“Thus the existence of a managerial hierarchy is a defining characteristic of the modern business
enterprise. A multiunit enterprise without such managers remains little more than a federation of
autonomous offices. [...] Such federations were often able to bring small reductions in information
and transactions costs but they could not lower costs through increased productivity. They could
not provide the administrative coordination that became the central function of modern business
enterprise.” 1

Consistent with this view, many firms respond to an increased need for coordination by abandoning
their decentralized structures and moving towards centralization.2 There are, however, also numer-
ous managers who argue that efficient coordination can be achieved in decentralized organizations
provided that division managers are able to communicate with each other. Alfred Sloan, the long-
time President and Chairman of General Motors, for instance, organized GM as a multi-divisional
firm and granted vast authority to the division managers.3,4 To ensure coordination between them,
Sloan set up various committees that gave the division managers an opportunity to exchange ideas.
1
Chandler (1977), pp.7-8.
2
See for instance the DaimlerChrysler Commercial Vehicles Division (Hannan, Podolny and Roberts 1999), Procter
& Gamble (Bartlett 1989) and Jacobs Suchard (Eccles and Holland 1989).
3
Our discussion of General Motors is based on “Co-ordination By Committee,” Chapter 7 in My Years With
General Motors by Alfred Sloan (1964).
4
Writing to some of his fellow GM executives in 1923, for instance, Sloan stated that “According to General Motors
plan of organization, to which I believe we all heartily subscribe, the activities of any specific Operation are under
the absolute control of the General Manager of that Division, subject only to very broad contact with the general
officers of the Corporation.” (Sloan 1964, p.106).

1
These committees did not diminish the authority of division managers to run their own divisions
and instead merely provided “a place to bring these men together under amicable circumstances
for the exchange of information and the ironing out of differences.” 5 Reflecting on the first such
committee, the General Purchasing Committee that was created to coordinate the purchasing of
inputs by the different divisions, Sloan wrote in 1924 that:

“The General Purchasing Committee has, I believe, shown the way and has demonstrated that
those responsible for each functional activity can work together to their own profit and to the profit of
the stockholders at the same time and such a plan of coordination is far better from every standpoint
than trying to inject it into the operations from some central activity.”6

The apparent success of Alfred Sloan’s GM and other decentralized firms in realizing inter-divisional
synergies suggests that in many cases coordination can indeed be achieved without centralization.7
The aim of this paper is to reconcile these conflicting views by analyzing when coordination
does and does not require centralization. Decentralized organizations have a natural advantage at
adapting decisions to local conditions since the decisions are made by the managers with the best
information about those conditions. However, such organizations also have a natural disadvantage
at coordinating since the manager in charge of one decision is uncertain about the decisions made
by others. Moreover, self-interested division managers may not internalize how their decisions
affect other divisions. One might therefore reason naively that centralization is optimal whenever
coordination is sufficiently important relative to the desire for responsiveness. We argue that this
reasoning is flawed and show that decentralization can be optimal even when coordination is very
important. Intuitively, when coordination becomes very important, division managers recognize
their interdependence and communicate and coordinate very well under decentralization. In con-
trast, under centralization, an increased need for coordination strains communication, as division
managers anticipate that headquarters will enforce a compromise. As a result, decentralization can
be optimal even when coordination becomes very important. In particular, this is so if the division
5
Sloan (1964), p.105.
6
Sloan (1964), p.104.
7
Other, and more recent, examples of firms that rely on the managers of largely autonomous divisions to coordi-
nate their activities without central intervention are PepsiCo (Montgomery and Magnani 2001) and AES Corporation
(Pfeffer 2004). In the 1980s and early 90s PepsiCo centralized very few of the activities of its three restaurant chains
Pizza Hut, Taco Bell and KFC and ran them as what were essentially stand-alone businesses. The management
of PepsiCo believed that coordination between the restaurant chains could be achieved by encouraging the divi-
sion managers to share information and letting them decide themselves on their joint undertakings: “In discussing
coordination across restaurant chains, senior corporate executives stressed that joint activity should be initiated by
divisions, not headquarters. Division presidents should have the prerogative to decide whether or not a given division
would participate in any specific joint activity. As one explained, ‘Let them sort it out. Eventually, they will. It
will make sense. They will get to the right decisions.’” (Montgomery and Magnani 2001, p.12).

2
managers’ incentives are sufficiently aligned and the divisions are not too different from each other.
Thus, only in organizations in which divisions are quite heterogenous, for instance in their relative
size, or in which implicit and explicit incentives strongly bias division managers towards their own
divisions, does coordination require centralization.
We propose a simple model of a multi-divisional organization with three main features: (i.)
decision making involves a trade-off between coordination and adaptation. In particular, two
decisions have to be made and the decision makers must balance the benefit of setting the decisions
close to each other with that of setting each decision close to its idiosyncratic environment or ‘state.’
Multinational enterprises (MNEs), for instance, may realize scale economies by coordinating the
product designs in different regions. These cost savings, however, must be traded off against the
revenue losses that arise when products are less tailored to local tastes. (ii.) information about the
states is dispersed and held by division managers who are biased towards maximizing the profits of
their own divisions rather than those of the overall organization. Moreover, the division managers
communicate their information strategically to influence decision making in their favor. In the
above MNE example, regional managers are likely to be best informed about the local tastes of
consumers and thus about the expected revenue losses due to standardization. In communicating
this information they have an incentive to behave strategically to influence the decision making to
their advantage. (iii.) the organization lacks commitment and, in particular, can only commit
to the ex ante allocation of decision rights (Grossman and Hart 1986 and Hart and Moore 1990).
This implies that decision makers are not able to commit to make their decisions dependent on
the information they receive in different ways. Communication therefore takes the form of cheap
talk (Crawford and Sobel 1982). In this setting we compare the performance of two organizational
structures: under Decentralization the division managers communicate with each other horizontally
and then make their decisions in a decentralized manner while under Centralization the division
managers communicate vertically with a headquarters manager who then makes both decisions.
Underlying our results are differences in how centralized and decentralized organizations aggre-
gate dispersed information. In our model vertical communication is always more informative than
horizontal communication. Essentially, since division managers are biased towards the profits of
their own divisions while headquarters aims to maximize overall profits, the preferences of a division
manager are more closely aligned with those of headquarters than with those of another division
manager. As a result, division managers share more information with headquarters than they do
with each other. Thus, while division managers are privately informed about their own divisions’
local conditions and therefore have an information advantage, headquarters has a communication

3
advantage. This communication advantage, however, gradually disappears as coordination be-
comes more important. In particular, whereas an increased need for coordination leads to worse
communication under Centralization, it actually improves communication under Decentralization.
Intuitively, when coordination becomes more important, headquarters increasingly ignores the in-
formation that it receives from the division managers about their local conditions. This induces
each manager to exaggerate his case more which, in turn, leads to less information being commu-
nicated. In contrast, under Decentralization an increase in the need for coordination makes the
managers more willing to listen to each other to avoid the costly coordination failures. As a result,
the managers’ incentives to exaggerate are mitigated and more information is communicated.
The fact that as coordination becomes more important the communication advantage of head-
quarters gradually disappears, while the division managers’ information advantage does not, drives
our central result: Decentralization can dominate Centralization even when coordination is ex-
tremely important. Specifically, we show that no matter how important coordination is, centraliz-
ing destroys value if the division managers are not too biased towards their own divisions and the
divisions are not too different from each other. Thus, coordination only requires centralization if
the divisions are quite heterogenous or division managers care mostly about their own divisions.
In the next section we discuss the related literature. In Section 3 we present our model which we
then solve in Sections 4 and 5. The main comparison between Centralization and Decentralization
is discussed in Section 6. In Section 7 we extend our model by allowing for asymmetries between
the divisions. Finally, we conclude in Section 8.

2 Related Literature

Our paper is related to, and borrows from, a few different literatures.

Coordination in Organizations: A number of recent papers analyze coordination in organizations.

Hart and Holmstrom (2002) focus on the trade-off between coordination and the private benefits
of doing things ‘independently:’ under decentralization the division managers do not fully internal-
ize the benefits of coordination while under centralization the decision maker ignores the private
benefits that division managers realize if they act independently. In Hart and Moore (2005), some
agents specialize in developing ideas about the independent use of one particular asset, while others
think about the coordinated use of several assets. They analyze the optimal hierarchical structure
and provide conditions under which coordinators should be superior to specialists. Finally, in Des-
sein, Garicano and Gertner (2005), ‘product managers’ are privately informed about the benefits of
running a particular division independently, whereas a ‘functional manager’ is privately informed

4
about the value of a coordinated approach. They endogenize the incentives for effort provision and
the communication of this private information and show that functional authority is preferred when
effort incentives are less important.
A key difference between these papers and ours is that in their models a trade-off between
centralization and decentralization arises because the incentives of the central decision-maker are
distorted (and biased towards coordination). In contrast, in our paper, authority is allocated to a
benevolent principal under centralization.8 Decentralization may nevertheless be strictly preferred
because it allows for a better use of dispersed information. In the context of the organization of
the overall economy, rather than a single firm, Bolton and Farrell (1990) have also emphasized this
trade-off between coordination and the use of local information. In a model of entry they show
that decentralization is good at selecting a low cost entrant but also results in inefficient delay
and duplication of entry. Unlike our paper, Bolton and Farrell (1990) rule out communication. In
contrast to our result, they find that centralization is preferred provided that coordination costs
are sufficiently large relative to the value of private information.9
Following Stein (1997), there is finally a large literature on internal capital markets which argues
that centralization can facilitate the efficient allocation of capital across otherwise independent
divisions. Closest in spirit to our paper are Stein (2002) and Friebel and Raith (2006) who assume
that the information necessary for an efficient allocation is dispersed and needs to be aggregated.10
Unlike in our paper, communication is always vertical.

Information Processing in Organizations: The large literature on team theory, starting with Mar-
shak and Radner (1972), constitutes the first attempt by economists to understand decision making
within firms. Team theory analyzes decision making in firms in which information is dispersed and
physical constraints make it costly to communicate or process this information. In doing so it
abstracts from incentive problems and assumes that agents act in the interest of the organization.
A team-theoretic model that is closely related to ours in spirit is Aoki (1986) who also compares
the efficiency of vertical and horizontal information structures. In contrast to this paper, and to
team theory in general, our results do not depend on assumptions about physical communication
constraints.11 Instead, we endogenize communication quality as a function of incentive conflicts. As
8
Otherwise, the trade-off between coordination and adaptation is similar to the trade-offs between coordination
and ‘independence’ considered in the above papers. Whereas in the above papers coordination is a binary choice, we
allow for decisions to be more or less coordinated.
9
Also in Gertner (1999) headquarters is unbiased but it only intervenes when bargaining between divisions breaks
down. His model further differs from ours in that sharing information is costly. He shows how the presence of an
independent arbitrator may foster information sharing.
10
See also Ozbas (2005) and Wulf (2005).
11
In addition to Aoki (1986), we follow Dessein and Santos (2006) in modeling a trade-off between adaptation and

5
such our analysis is related to the mechanism design approach to organizational design which also
focuses on the incentives of agents to misrepresent their information.12 This literature, however,
concentrates on settings in which the Revelation Principle holds and in which centralized organi-
zations are therefore always weakly optimal. In contrast, we develop a simple model in which the
Revelation Principle does not hold since agents are unable to commit to mechanisms. As a result
decentralized organizations can be strictly optimal.
Our no-commitment assumption is in line with a number of recent papers that adopt an incom-
plete contracting approach to organizational design and model communication as a cheap talk.13
Dessein (2002) considers a model in which a principal must decide between delegating decision
rights to an agent versus keeping control and communicating with that agent.14 Harris and Raviv
(2005) consider a similar set up but allow the principal to have private information while Alonso
and Matouschek (2006) endogenize the commitment power of the principal in an infinitely repeated
game. These papers, however, do not analyze coordination, nor do they allow for horizontal com-
munication. A paper that does analyze coordination in a setting with horizontal communication is
Rantakari (2006). He has independently developed a model that is similar to ours and shows that
if division managers care about coordination to different degrees, then it may be optimal to cen-
tralize control in the division that cares less about coordination. Since in his model, managers care
exclusively about their own division, centralization is always optimal when coordination becomes
important.15

CheapTalk: From a methodological perspective our paper is related to the large cheap talk literature
that builds on Crawford and Sobel (1982). A technical difference between our model and that in
coordination. This paper shows how, in the presence of imperfect communication, extensive specialization results in
organizations that ignore local knowledge. Also Cremer, Garicano and Prat (2006) study how physical communication
constraints limit coordination. In their model, organizations face a trade-off between adopting a common technical
language, which allows for better coordination among units, or several specialized, distinct languages which are better
adapted to each unit. Other related team theory papers include Bolton and Dewatripont (1994), on the firm as a
communication network, Garicano (2000), on the organization of knowledge in production, Qian, Roland and Xu
(2006), on coordinating change in organizations, and more generally, the large literature on information processing
organizations (Radner 1993 and others; see Van Zandt 1999 for an overview).
12
For a survey of this literature see Mookherjee (2006).
13
Another related literature analyzes how adding a prior cheap talk stage matters in coordination games or games
with asymmetric information. Farrell (1987), for example, shows how cheap talk can improve coordination in a battle
of the sexes game. More closely related to our set-up, Farrell and Gibbons (1989) analyzes how cheap talk can avoid
a break-down in a bargaining game with asymmetric information. A recent paper along these lines is Baliga and
Morris (2002). Farrell and Rabin (1996) provide an overview of related cheap talk applications.
14
See also Marino and Matsusaka (2005).
15
Our paper is also related to the large political economy literature on fiscal federalism which studies the choice
between centralization and decentralization in the organization of states (see Oates 1999 and Lockwood 2005 for
surveys). However, the institutions they look at, such as elections and legislatures, are different from the institutions
that determine decision making within firms and that we focus on.

6
Crawford and Sobel (1982) is that we allow for the preferred decisions of the senders and receivers
to coincide. As such our paper is related to Melumad and Shibano (1990) who also allow for this
possibility. A key difference between their analysis and ours is that they focus on communication
equilibria with a finite number of intervals while we allow for equilibria with an infinite number
of intervals. We show that such equilibria maximize the expected joint surplus and that they are
computationally straightforward since they avoid the integer problems associated with finite interval
equilibria. For this reason we believe that our model is more tractable than the leading example in
Crawford and Sobel, the traditional workhorse for cheap talk applications.16 We further differ from
both Crawford and Sobel (1982) and Melumad and Shibano (1990) in that we allow for multiple
senders. Our paper is therefore related to Battaglini (2002) and Krishna and Morgan (2002) who
consider models in which a principal consults multiple informed “experts” who all observe the same
piece of information. The question they investigate is whether, and if so how, the principal can
illicit this information from the experts. In contrast, in our model the senders observe different
pieces of independent information which makes it impossible to achieve truth-telling.

3 The Model

An organization consists of two operating divisions, Division 1 and Division 2, and potentially one
headquarters. Division j ∈ {1, 2} generates profits that depend on its local conditions, described
by θj ∈ R, and on two decisions, d1 ∈ R and d2 ∈ R. In particular, the profits of Division 1 are
given by
π 1 = K1 − (d1 − θ1 )2 − δ (d1 − d2 )2 , (1)

where K1 ∈ R+ is the maximum profit that the division can realize. The first squared term
captures the adaptation loss that Division 1 incurs if decision d1 is not perfectly adapted to its
local conditions, that is if d1 6= θ1 , and the second squared term captures the coordination loss
that Division 1 incurs if the two decisions are not perfectly coordinated, that is if d1 6= d2 . The
parameter δ ∈ [0, ∞) then measures the importance of coordination relative to adaptation. The
profits of Division 2 are similarly given by

π 2 = K2 − (d2 − θ2 )2 − δ (d1 − d2 )2 , (2)

where K2 ∈ R+ is the maximum profit that Division 2 can realize. Without loss of generality we
set K1 = K2 = 0. Headquarters does not generate any profits.
16
A recent paper which shares this feature is Kawamura (2006).

7
Information: Each division is run by one manager. Manager 1, the manager in charge of Division
1, privately observes his local conditions θ1 but does not know the realization of θ2 . Similarly,
Manager 2 observes θ2 but does not know θ1 . The HQ Manager, i.e. the manager in charge of
headquarters, observes neither θ1 nor θ2 . It is common knowledge, however, that θ1 and θ2 are
uniformly distributed on [−s1 , s1 ] and [−s2 , s2 ] respectively, with s1 and s2 ∈ R+ . The draws of θ1
and θ2 are independent.
Preferences: We assume that Manager 1 maximizes λπ 1 + (1 − λ)π 2 whereas Manager 2 maximizes
(1−λ)π 1 +λπ 2 , where λ ∈ [1/2, 1] . The parameter λ thus captures how biased each division manager
is towards his own division’s profits. The HQ Manager simply aggregates the preferences of the two
division managers and thus maximizes by π 1 + π 2 .17 For simplicity we take the preferences of the
managers as given and do not model their origins. Intuitively, factors outside of our model, such
as career concerns and subjective performance evaluations, are prone to bias division managers
towards maximizing the profits of the division under their direct control as their managerial skills
and effort will be mainly reflected in the performance of this division. In contrast, the skills
and effort of the HQ Manager are more likely to be reflected in the overall performance of the
organization, rather than in that of one particular division. In principle, the organization might
attempt to neutralize the division managers’ biases towards their own divisions by compensating
them more for the performance of the rest of the organization than for that of their own division.
As will become clear below, if it were possible to contract over λ, the organization would always set
λ = 1/2 and all organizational structures would perform equally well. However, it will typically be
undesirable for the organization to fully align managerial incentives if the division managers have
to make division specific effort choices.18 Furthermore, to the extent that divisions need to make
many decisions, the allocation of one particular decision right is likely to have only a negligible
impact on endogenously derived incentives. It therefore seems reasonable, as a first step, to assume
that the division managers’ biases do not differ across organizational structures.19
17
For the results presented below it is not important that the HQ Manager is entirely unbiased. Qualitatively
similar results would be obtained as long as her utility function is a convex combination of that of Managers 1 and
2.
18
The conflict between motivating efficient effort provision, on the one hand, and efficient decision making and/or
communication, on the other, has been analyzed in a number of recent papers (Athey and Roberts 2001; Dessein,
Garicano and Gertner 2005; Friebel and Raith 2006). These papers show that it is typically optimal for organizations
to bias division managers towards their own divisions to motivate effort provision even if doing so distorts their
incentives on other dimensions.
19
In reality λ can be very big and in some cases it can even be equal to one. GM provides a historical example of
such a case: “Under the incentive system in operation before 1918, a small number of division managers had contracts
providing them with a stated share in the profits of their own divisions, irrespective of how much the corporation as a
whole earned. Inevitably, this system exaggerated the self-interest of each division at the expense of the interests of
the corporation itself. It was even possible for a division manager to act contrary to the interests of the corporation

8
Contracts and Communication: We follow the property rights literature (Grossman and Hart 1986,
Hart and Moore 1990) in assuming that contracts are highly incomplete. In particular, the orga-
nization can only commit to an ex ante allocation of decision rights. Agents are unable to contract
over the decisions themselves and over the communication protocol that is used to aggregate in-
formation. Once the decision rights have been allocated, they cannot be transferred before the
decisions are made. We focus on two allocations of decision rights. Under Decentralization Man-
ager 1 has the right to make decision d1 and Manager 2 has the right to make decision d2 and both
decisions are made simultaneously. Under Centralization both decision rights are held by the HQ
Manager.20
The lack of commitment implies that the decision makers are not able to commit to paying
transfers that depend on the information they receive or to make their decisions depend on such
information in different ways. Communication therefore takes the form of an informal mechanism:
cheap talk. For simplicity we assume that this informal communication occurs in one round of com-
munication. In particular, under Decentralization Manager 1 sends message m1 ∈ M1 to Manager
2 and, simultaneously, Manager 2 sends message m2 ∈ M2 to Manager 1. Under Centralization,
Managers 1 and 2 simultaneously send messages m1 ∈ M1 and m2 ∈ M2 to the HQ Manager.
We refer to communication between the division managers as horizontal communication and that
between the division managers and the HQ Manager as vertical communication. It is well known
in the literature on cheap talk games that repeated rounds of communication can expand the set
of equilibrium outcomes even if only one player is informed.21 However, even for a simple cheap
talk game such as the leading example in Crawford and Sobel (1982), it is still an open question
as to what is the optimal communication protocol. Since it is our view that communication is an
‘informal’ mechanism which cannot be structured by the mechanism designer, it seems reasonable
to focus on the simplest form of informal communication. In this sense, we take a similar approach
as the property rights literature which assumes that players engage in ex post bargaining but limits
the power of the mechanism designer to structure this bargaining game.
A key feature of our model is its symmetry: the divisions are of equal size, they have the same
need for coordination and the two decisions are made simultaneously. We focus on a symmetric
organization since it greatly simplifies the analysis. In Section 7, however, we allow for asymmetries
between the divisions and discuss how such asymmetries affect our results.
in his effort to maximize his own division’s profits.” (Sloan 1964, p.409).
20
A natural variation of Decentralization is to allow for sequential decision making and a natural variation of
Centralization is to centralize both decision rights in one of the divisions. It turns out that the former structure
dominates the latter. We discuss the former structure in Section 7.
21
See, for example, Aumann and Hart (2003) and Krishna and Morgan (2004).

9
 t1 t2 t3 t4

Decision rights Managers 1 and 2 Managers 1 and 2 Decisions d1 & d2

are allocated learn states θ1 & send messages m1 & are made.
θ 2 respectively. m2 respectively.

Figure 1: Timeline

The game is summarized in Figure 1. First, decision rights are allocated to maximize the total
expected profits E [π 1 + π 2 ]. Under Centralization the HQ Manager gets the right to make both
decisions and under Decentralization each division manager gets the right to make one decision.
Second, the division managers become informed about their local conditions, i.e. they learn θ1
and θ2 respectively. Third, the division managers communicate with the decision makers. Under
Centralization they engage in vertical communication, sending messages m1 and m2 to the HQ
Manager, while under Decentralization they engage in horizontal communication, sending messages
m1 and m2 to each other. Finally, the decisions d1 and d2 are made. Each decision maker chooses
the decision that maximizes his or her payoff given the information that has been communicated.

4 Decision Making

In this section we characterize decision making under Centralization and Decentralization, taking
as given the posterior beliefs of θ1 and θ2 . In Section 5 we then characterize the communication
subgame and hence endogenize these beliefs. Finally, in Section 6 we draw on our understanding
of the decision making and the communication subgame to compare the performance of the two
organizational structures. The proofs of all lemmas and propositions are in the appendix.
Under Centralization, the HQ Manager receives messages m1 and m2 from the division managers
and then chooses the decisions d1 and d2 that maximize E [π 1 + π 2 | m], where π 1 and π 2 are the
divisions’ profits as given by (1) and (2), respectively, and m ≡ (m1 , m2 ). The decisions that solve
this problem are convex combinations of the HQ Manager’s posterior beliefs of θ1 and θ2 given m:

dC
1 ≡ γ C EH [θ 1 | m] + (1 − γ C ) EH [θ 2 | m] (3)

and
dC
2 ≡ (1 − γ C ) EH [θ 1 | m] + γ C EH [θ 2 | m] , (4)

10
where
1 + 2δ
γC ≡ (5)
1 + 4δ
and EH [θj | m], j = 1, 2, is the HQ Manager’s expected value of θj after receiving messages m.
Note that γ C is decreasing in δ and ranges from 1/2 to 1. When the decisions are independent,
that is when δ = 0, the HQ Manager sets dC C
1 = E [θ 1 | m] and d2 = E [θ 2 | m]. As the importance
of coordination δ increases, she puts less weight on E [θ1 | m] and more weight on E [θ2 | m] when
making decision d1 . Eventually, as δ → ∞, she puts the same weight on both decisions, i.e. she
sets dC C
1 = d2 = E [θ 1 + θ 2 | m] /2.
Under Decentralization, the division managers first send each other messages m1 and m2 . Once
the messages have been exchanged, Manager 1 chooses d1 to maximize E [λπ 1 + (1 − λ)π 2 | m] and,
simultaneously, Manager 2 chooses d2 to maximize E [(1 − λ)π 1 + λπ 2 | m]. The decision that
Manager 1 makes is a convex combination of his local conditions, θ1 , and of E1 [d2 | m] , i.e. the
decision that Manager 1 expects Manager 2 to make:
λ δ
d1 = θ1 + E1 [d2 | m] . (6)
λ+δ λ+δ
Similarly, we have that
λ δ
d2 = θ2 + E2 [d1 | m] , (7)
λ+δ λ+δ
where E2 [d1 | m] is the decision that Manager 2 expects Manager 1 to make. It is intuitive that
the weight that each division manager puts on his state is increasing in the own-division bias λ and
decreasing in the importance of coordination δ.
The decisions E1 [d2 | m] and E2 [d1 | m] that each division manager expects his counterpart to
make can be obtained by taking the expectations of (6) and (7). Doing so and substituting back
into (6) and (7) we find that Manager 1’s decision is a convex combination of his local conditions
θ1 , his posterior belief about θ2 and Manager 2’s posterior belief about θ1 :
µ ¶
D λ δ δ λ+δ
d1 ≡ θ1 + E2 [θ1 | m] + E2 [θ2 | m] . (8)
λ+δ λ + δ λ + 2δ λ + 2δ
Similarly, we have that
µ ¶
λ δ λ+δ δ
dD
2 ≡ θ2 + E2 [θ1 | m] + E2 [θ2 | m] . (9)
λ+δ λ+δ λ + 2δ λ + 2δ
It can be seen that as δ increases, Manager j = 1, 2 puts less weight on his private information
θj and more weight on a weighted average of the posterior beliefs. In the limit, as δ → ∞, the
division managers only rely on the communicated information and set dD D
1 = d2 = (E2 [θ 1 | m] +
E2 [θ2 | m])/2. Note that, for given posteriors, these are exactly the same decisions that the HQ
Manager implements when δ → ∞.

11
5 Strategic Communication

In this section we analyze communication under the two organizational structures. A key insight of
this section is that while vertical communication is more informative than horizontal communication
this difference decreases, and eventually vanishes, as coordination becomes more important. This
is so since an increase in the need for coordination improves horizontal communication but worsens
vertical communication.
We proceed by investigating the division managers’ incentives to misrepresent information in
the next sub-section. This then allows us to characterize the communication equilibria in Section
5.2 and to compare the quality of vertical and horizontal communication in Section 5.3. Finally, in
Section 5.4 we derive the expected profits and show that they can be expressed as linear functions
of the quality of communication.

5.1 Incentives to Misrepresent Information

Regardless of the allocation of control, managers have an incentive to make others believe that
their local conditions are more extreme, that is, further from the average than is truly the case.
To see this, consider first the incentives of Manager 1 to misrepresent his information under Cen-
tralization.22 For this purpose let ν 1 = EH [θ1 | m1 ] be the HQ Manager’s expectation of θ1 after
receiving message m1 and suppose that Manager 1 can simply choose any ν 1 . In other words,
suppose that Manager 1 can credibly misrepresent his information about his state. The question
then is why, and by how much, Manager 1 would like to misrepresent his information. Clearly,
Manager 1 would like the HQ Manager to have the posterior that maximizes his expected payoff:
h i
ν ∗1 = arg max E −λ (d1 − θ1 )2 − (1 − λ) (d2 − θ2 )2 − δ (d1 − d2 )2 | θ1 , (10)
ν1

where d1 = dC C
1 and d2 = d2 as defined in (3) and (4) respectively. We will see below that in
equilibrium the expected value of the posterior of θ2 is equal to the expected value of θ2 , i.e. that
Em2 [E [θ2 |m2 ]] = E [θ2 ] = 0. Assuming that this relationship holds we can use (10) to obtain
(2λ − 1) δ
ν ∗1 − θ1 = θ1 ≡ bC θ1 . (11)
λ+δ
This expression shows that Manager 1 has no incentive to misrepresent his information when θ1 = 0.
The reason for this is that, in the absence of any information about θ2 , the incentives of Manager
1 and the HQ Manager are perfectly aligned when θ1 = 0. For θ1 6= 0, however, Manager 1 does
have an incentive to exaggerate his state, that is to set |ν ∗1 − θ1 | > 0. This is due to his desire to
22
The argument for Manager 2 is analogous.

12
adapt d1 more closely to θ1 . Essentially, since Manager 1 does not fully internalize the need to
coordinate the decisions, he always wants d1 to be set closer to θ1 than the HQ Manager does. It
is therefore in his interest to induce the HQ Manager to choose a larger decision whenever θ1 > 0
and to choose a smaller decision whenever θ1 < 0 and he can achieve this by exaggerating his
state. Expression (11) also shows Manager 1’s desire to exaggerate his state is increasing in |θ1 |.
Essentially, the bigger |θ1 |, the bigger Manager 1’s marginal benefit of inducing the HQ Manager
to implement a more extreme decision and thereby bring d1 closer to θ1 .
The division managers’ incentives to misrepresent information under Decentralization are similar
to those under Centralization. To see this, let ν 1 = E2 [θ1 | m1 ] be Manager 2’s expectation of
θ1 after receiving message m1 and suppose again that Manager 1 can simply choose any ν 1 . His
optimal choice of ν 1 is given by (10), where d1 = dD D
1 and d2 = d2 as defined in (8) and (9)
respectively. If we assume again that Em2 [E [θ2 |m2 ]] = E [θ2 ] = 0, which will be shown to hold in
equilibrium, then it follows that for δ > 0
(2λ − 1) (λ + δ)
ν ∗1 − θ1 = θ1 ≡ bD θ1 . (12)
λ (1 − λ) + δ
Thus, it is again the case that Manager 1 has no incentive to misrepresent his information when
θ1 = 0, that he has an incentive to exaggerate it if θ1 6= 0 and that his incentive to exaggerate is
increasing in |θ1 |. The incentive to exaggerate is again due to Manager 1’s desire to closely adapt
decision d1 to θ1 . In particular, since Manager 1 puts some weight on coordinating his decision
with that of Manager 2, his ability to adapt d1 to θ1 is constrained by Manager 2’s choice of d2 .
If, for instance, θ1 > 0, then Manager 1 would like to induce Manager 2 to set a higher d2 so that
he can increase d1 and thereby bring it closer to θ1 . He can do so by exaggerating his state, i.e.
by setting ν ∗1 − θ1 > 0.
How do the incentives to misrepresent information depend on the own-division bias and the
need for coordination? It is intuitive that under both structures the incentives to misrepresent
are increasing in the own-division bias, i.e. dbl /dλ ≥ 0 for l = C, D. What might be less
intuitive is that an increase in the need for coordination increases the incentives to misrepresent
information under Centralization but decreases them under Decentralization, i.e. dbC /dδ ≥ 0 and
dbD /dδ ≤ 0. This suggests that, as coordination becomes more important, the division managers
are less willing to share information with the coordinator under Centralization but are more willing
to share information with each other under Decentralization. To understand this, recall that under
Centralization Manager 1 exaggerates his private information to induce the HQ Manager to make a
more extreme decision, i.e. to increase |d1 |, while under Decentralization he exaggerates to induce
Manager 2 to make a more extreme decision. The key observation is that the more important the

13
need for coordination, the less sensitive the HQ Manager’s decision making is to the information
she receives from Manager 1 and thus the stronger Manager 1’s desire to exaggerate. In contrast,
an increase in the importance of coordination makes Manager 2’s decision more sensitive to the
information he receives from Manager 1 and thereby reduces Manager 1’s desire to misrepresent his
information. Essentially, when coordination becomes more important, division managers are less
willing to share information with a headquarters that increasingly ignores their information but
they are more willing to share information with each other to avoid costly coordination failures.23

5.2 Communication Equilibria

We now show that, as in Crawford and Sobel (1982), all communication equilibria are interval
equilibria in which the state spaces [−s1 , s1 ] and [−s2 , s2 ] are partitioned into intervals and the
division manager only reveal which interval their local conditions θ1 and θ2 belong to. In this sense
the managers’ communication is noisy and information is lost. Moreover, the size of the intervals —
which determines how noisy communication is — depends directly on bD and bC as defined in (11)
and (12).
A communication equilibrium under each organizational structure is characterized by (i.) com-
munication rules for the division managers, (ii.) decision rules for the decision makers and (iii.)
belief functions for the message receivers. The communication rule for Manager j = 1, 2 specifies
the probability of sending message mj ∈ Mj conditional on observing state θj and we denote it by
μj (mj | θj ). The decision rules map messages m1 ∈ M1 and m2 ∈ M2 into decisions d1 ∈ R and
d2 ∈ R and we denote them by d1 (m) and d2 (m) respectively. Finally, the belief functions are
denoted by gj (θj | mj ) for j = 1, 2 and state the probability of state θj conditional on observing
message mj .
We focus on Perfect Bayesian Equilibria of the communication subgame which require that
(i.) communication rules are optimal for the division managers given the decision rules, (ii.) the
decision rules are optimal for the decision makers given the belief functions and (iii.) the belief
functions are derived from the communication rules using Bayes’ rule whenever possible.
Since all communication equilibria will be shown to be interval equilibria, we denote by a2N
j ≡
23
To see this formally,
  note that before Manager 1 learns Manager 2’s message, he expects the HQ Manager to
make decision Eθ2 dC 1 = γ C ν 1 , where γ C is given in (5) and ν 1 = EH (θ1 | m1 ) is the HQ Manager’s posterior belief
of θ1 . Since γ C is decreasing in δ, the HQ Manager’s decision making becomes less sensitive to the information she
receives from Manager 1 when coordination becomes more important. This, in turn, gives Manager 1 an incentive to
exaggerate his state more. In contrast, we can use (9) to show that under Decentralization the  decision
 that Manager
1 expects Manager 2 to make before having received Manager 2’s message is given by Eθ2 dD 2 = ν 1 δ/(λ + 2δ). It
 
can be seen that Eθ2 dD 2 is more sensitive to ν 1 when coordination is more important. As a result, an increase in
δ actually mitigates Manager 1’s desire to exaggerate θ1 .

14
(aj,−N , ..., aj,−1 , aj,0 , aj,1 , ..., aj,N ) and a2N−1
j ≡ (aj,−N , ..., aj,−1 , aj,1 , ..., aj,N ) the partitioning of
[−sj , sj ] into 2N and 2N − 1 intervals respectively, where aj,−N = −sj , aj,0 = 0 and aj,N = sj .
Thus, a2N
j corresponds to finite interval equilibria with an even number of intervals and a2N−1
j
corresponds to those with an odd number of intervals. As will be shown in the next proposition,
the end points are symmetrically distributed around zero, i.e. aj,i = aj,−i for all i ∈ {1, ..., N }.
The following proposition characterizes the finite communication equilibria when δ > 0.

PROPOSITION 1 (Communication Equilibria). If δ ∈ (0, ∞), then for every positive integer Nj ,
j = 1, 2, there exists at least one equilibrium (μ1 (·), μ2 (·) , d1 (·), d2 (·), g1 (·), g2 (·)), where

i. μj (mj | θj ) is uniform, supported on [aj,i−1 , aj,i ] if θj ∈ (aj,i−1 , aj,i ),

ii. gj (θj | mj ) is uniform supported on [aj,i−1 , aj,i ] if mj ∈ (aj,i−1 , aj,i ),

iii. aj,i+1 − aj,i = aj,i − aj,i−1 + 4baj,i for i = 1, ..., Nj − 1

aj,−(i+1) − aj,−i = aj,−i − aj,−(i−1) + 4baj,−i for i = 1, ..., Nj − 1

with b = bC under Centralization, where bC is defined in (11), and

b = bD under Decentralization, where bD is defined in (12),

iv. dj (m) = dC
j , j = 1, 2, under Centralization, where dC
j are given by (3) and (4) and

dj (m) = dD D
j , j = 1, 2, under Decentralization, where dj are given by (8) and (9).

Moreover, all other finite equilibria have relationships between θ1 and θ2 and the managers’ choices
of d1 and d2 that are the same as those in this class for some value of N1 and N2 ; they are
therefore economically equivalent.

The communication equilibria are illustrated in Figure 2. In these equilibria each division
manager communicates what interval his state lies in. The size of the intervals is determined by
the difference equations in Part (iii.) of the proposition. The size of an interval (aj,i+1 − aj,i ) equals
the size of the preceding interval (aj,i − aj,i−1 ), plus respectively 4bC aj,i under Centralization and
4bD aj,i under Decentralization, where aj,i is the dividing point between the two intervals. Recall
from Section 5.1, that bC θj , j = 1, 2, is the difference between the true state of nature θj and
what Manager j would like the HQ Manager to believe that the state is. Similarly, bD θj , j = 1, 2,
represents by how much Manager j wants to misrepresent his state when talking to the other division
manager under Decentralization. The incentives to distort information thus directly determine how
quickly communication deteriorates as θj is further away from its mean.24 It is intuitive that since
24
This can be related to the leading example in Crawford and Sobel (1982). In that model, there is a fixed difference
b between the the true state of nature and what the sender would like the receive to believe is the true state, and

15
 θ1

Division managers send

messages indicating the intervals
their states lie in

Decision makers update their

θ2 beliefs

Decision makers make

their decisions

Figure 2: Communication Equilibria

the incentives to misrepresent information are increasing in |θj |, not only is it the case that less
information is transmitted the larger |θj | , but also the rate at which communication becomes noisier
is increasing in |θj | .
Proposition 1 characterizes communication equilibria for δ > 0. In the absence of any need for
coordination the communication equilibria are straightforward. In particular, under Centraliza-
tion, truth-telling can be sustained if δ = 0 since, in this case, there is no incentive conflict between
the division managers. While there are other equilibria, we assume in the remaining analysis
that when δ = 0 the managers coordinate on the truth-telling equilibrium. Under Decentraliza-
tion, communication is irrelevant when δ = 0 since it is optimal for each division manager to set his
decision equal to his state which, of course, he observes directly. This implies that for δ = 0 all com-
munication strategies are consistent with a Perfect Bayesian Equilibrium under Decentralization.
Merely to facilitate the exposition of one comparative static that we perform later, and without
losing generality, we assume that for δ = 0 the communication equilibrium under Decentralization
is as those described in Proposition 1.25
Proposition 1 shows that there does not exist an upper limit on the number of intervals that can
equivalently, intervals grow at a fixed rate of 4b rather than 4baj,i .
25
See discussion of equation (17).

16
be sustained in equilibrium. This is in contrast to Crawford and Sobel (1982) where the maximum
number of intervals is always finite. This difference is due to the fact that in our model there
exists a state, namely θj = 0 for j = 1, 2, in which the incentives of the sender and the receiver
are perfectly aligned.26 In Crawford and Sobel (1982) this possibility is ruled out. The next
proposition shows that in the limit in which the number of intervals goes to infinity, the strategies
and beliefs described in Proposition 1 constitute a Perfect Bayesian Equilibrium and that the total
expected profits are maximized in this equilibrium.

PROPOSITION 2 (Efficiency). The limit of strategy profiles and beliefs (μ1 (·), μ2 (·), d1 (·), d2 (·),
g1 (·), g2 (·)) as N1 , N2 → ∞ is a Perfect Bayesian Equilibrium of the communication game. In
this equilibrium the total expected profits E [π 1 + π 2 ] are higher than in any other equilibrium.

An illustration of an equilibrium in which the number of intervals goes to infinity is provided in

Figure 2. In such an equilibrium the size of the intervals is infinitesimally small when θj , j = 1, 2,
is close to zero but grows as |θj | increases. We believe that it is reasonable to assume that the
organization is able to coordinate on the equilibrium that maximizes the expected overall profits
and we focus on this equilibrium for the rest of the analysis.

5.3 Communication Comparison

We can now compare the quality of communication under the two organizational structures and
analyze how it is affected by changes in the need for coordination and the own-division bias. We
h i
measure the quality of communication as the residual variance E (θj − E [θj |mj ])2 of θj , j = 1, 2.
The next lemma derives the residual variance under vertical and horizontal communication.

LEMMA 1. In the most efficient equilibrium in which N1 , N2 → ∞ the residual variance is given
by
h i
E (θj − E [θj |mj ])2 = Sl σ 2j j = 1, 2 and l = C, D,

where
bl
Sl = . (13)
3 + 4bl

The residual variance is therefore directly related to the division managers’ incentives to mis-
represent information as defined in (11) and (12). In particular, when bl = 0, l = C, D, then the
division managers perfectly reveal their information and as a result the residual variance is zero.
As bl increases, less information is communicated in equilibrium and the residual variance increases.
26
We share this feature with Melumad and Shibano (1991); see the related literature in Section 2.

17
 1/4

SD

SC

0
1/2 1 λ

Figure 3: Communication Comparison (λ)

Finally, as bl → ∞, the division managers only reveal whether their state is positive or negative
and thus Sl → 1/4. We can now state the following proposition.

PROPOSITION 3 (The Quality of Communication).

i. SC = SD = 0 if λ = 1/2 and SD > SC otherwise,
ii. ∂SD /∂λ > ∂SC /∂λ > 0,
iii. ∂SC /∂δ > 0 > ∂SD /∂δ and limδ→∞ SD = limδ→∞ SC .
Parts (i.) and (ii.), which are illustrated in Figure 3, show that vertical communication is
in general more efficient than horizontal communication. In other words, headquarters has a
communication advantage vis-a-vis the division managers. To understand this, recall that for
λ = 1/2 communication is perfect under both organizational structures and note, from Part (ii.),
that an increase in the own-division bias λ has a more detrimental effect on horizontal than on
vertical communication. This is the case since, under Centralization, an increase in λ increases
the bias of the senders but does not affect the decision making of the receiver. In contrast, under
Decentralization, an increase in λ also leads to more biased decision making by the receiver.
Part (iii.) is illustrated in Figure 4 and shows that the communication advantage diminishes as
the need for coordination increases. As discussed in Section 5.1 this key property is due to the fact
that a higher δ increases the incentives of division managers to misrepresent their information under
Centralization but reduces them under Decentralization. Part (iii.) also shows that as δ increases
the communication advantage not only shrinks but actually vanishes. This can be understood by

18
 1/4

SD

SC

0
0 1 δ
1+δ

Figure 4: Communication Comparison (δ)

recalling from Section 4 that in the limit in which δ → ∞ the decision making under Centralization
and under Decentralization converge: under both structures the decisions are set equal to the
average posterior. It is then not surprising that the quality of communication also converges.

5.4 Organizational Performance

We can now state the expected profits for each organizational structure.

PROPOSITION 4 (Organizational Performance). Under Centralization the expected profits are

given by
¡ ¢
ΠC = − (AC + (1 − AC ) SC ) σ 21 + σ 22 , (14)

and under Decentralization they are given by

¡ ¢
ΠD = − (AD + BD SD ) σ 21 + σ 22 , (15)

where ¡ ¢
2δ 2 λ2 + δ δ 4λ3 + 6λ2 δ + 2δ 2 − λ2
AC ≡ , AD ≡ and BD ≡ δ 2 . (16)
1 + 4δ (λ + 2δ)2 (λ + δ)2 (λ + 2δ)2

The proposition shows that under both organizational structures the expected profits are a linear
¡ ¢ ¡ ¢
function of the underlying uncertainty σ 21 + σ 22 and the sum of the residual variances Sl σ 21 + σ 22 ,
l = C, D.

19
 Since the HQ Manager’s decision making is efficient, the first best expected profits would be
realized if she were perfectly informed, i.e. if SC = 0. It then follows from (14) that the first
¡ ¢
best expected profits are given by −AC σ 21 + σ 22 . It is intuitive that these expected profits are
decreasing in the need for coordination and are independent of the division managers’ biases.
The expected profits under Centralization differ from the first best benchmark since the HQ
Manager is in general not perfectly informed. In particular, the more biased the division managers
are towards their own divisions, the less information is communicated and thus the lower the
expected profits. The expected profits are also decreasing in δ: not only does an increase in the
need for coordination lead to worse communication but it also reduces the expected profits for any
given communication quality.
The expected profits under Decentralization differ from the first best both because of imperfect
communication and because the division managers’ decision making is biased. Since an increase in
the own-division bias leads to worse communication and to more biased decision making it clearly
reduces the expected profits. The impact of an increase in δ, in contrast, is ambiguous: it improves
communication but it reduces the expected profits for any given communication quality.

6 Centralization versus Decentralization

We can now compare the performance of the two organizational structures. A clear advantage of
a decentralized organization is that it puts in control those managers who are closest to the local
information. In contrast, in a centralized organization some of this information is lost when it is
communicated to the decision maker. Naturally, the lack of local information impairs the ability
of a centralized organization to adapt decisions to the local conditions.
However, while Decentralization has an advantage at adapting decisions to local conditions, it
has a disadvantage at ensuring that the decisions are coordinated. This is so for two reasons. First,
the division managers do not fully internalize the need for coordination and, as such, put excessive
weight on adapting their decisions to the local conditions. Second, effective coordination requires
that the manager who makes one decision knows what the other decision is. Under Centralization
this is naturally the case since both decisions are made by the same manager. In contrast, under
Decentralization Manager j = 1, 2 is uncertain about what decision Manager k 6= j will make
since communication between them is imperfect. In sum, division managers lack both the right
incentives and the right information to ensure effective coordination while headquarters lacks the
local information to efficiently adapt decisions to the local conditions.
The next lemma shows that these factors translate into a coordination advantage of Cen-

20
tralization and an adaptation advantage of Decentralization. To state this lemma, let ALl =
h¡ ¢2 ¡ ¢2 i
E dl1 − θ1 + dl2 − θ2 , l = C, D, denote the adaptation losses under the two organizational
h¡ ¢2 i
structures and let CLl = E dl1 − dl2 denote the coordination losses.

LEMMA 2. For all λ ∈ [1/2, 1] and δ ≥ 0, ALC ≥ ALD and CLC ≤ CLD .

This lemma supports the basic intuition that centralized structures perform relatively well in
terms of coordination and decentralized structures perform relatively well in terms of adaptation.
From this one could reason naively that Centralization will prevail if the need for coordination is
sufficiently important. A key insight of this section is that this reasoning is flawed. To see this,
consider the next proposition which compares the performance of the two organizational structures
for λ > 1/2 and δ > 0. For λ = 1/2 or δ = 0 both organizational structures achieve the first best
expected profits since, in this case, there is no incentive conflict between the managers.

PROPOSITION 5 (Centralization versus Decentralization). Suppose that λ > 1/2 and δ > 0:
¡ ¢
i. δ small: if δ ∈ 0, δ , then Decentralization strictly dominates Centralization, where
δ > 0 is defined in the proof .
¡ ¢
ii. δ large: if δ ∈ δ, ∞ , then Decentralization strictly dominates Centralization for all
¡ ¢
λ ∈ 1/2, λ(δ) and Centralization strictly dominates Decentralization for all
¡ ¤
λ ∈ λ(δ), 1 , where λ(δ) > 17/28 is defined in the proof.
iii. δ → ∞: as δ → ∞, the expected profits under Decentralization and Centralization con-
verge, i.e. limδ→∞ ΠD = limδ→∞ ΠC .

The proposition is illustrated in Figure 5, where the downward sloping curve represents λ(δ).
Part (i.) shows that while organizational structure is irrelevant when δ = 0, Decentralization strictly
dominates Centralization if the need for coordination is small but positive and this for any own-
division bias λ > 1/2. Part (ii.) shows that in spite of the coordination advantage of Centralization,
Decentralization can strictly outperform Centralization no matter how important coordination is.
In particular, whenever λ < 17/28, Decentralization is optimal for any finite δ > 0. This also
implies that the Delegation Principle holds in our model. The Delegation Principle states that
decision rights should be delegated to the managers with the most information provided that their
incentives are sufficiently aligned and it is satisfied in our model since Decentralization strictly
dominates Centralization for sufficiently small values of λ > 1/2.27 Finally, Part (iii.) shows
that in the limit in which coordination becomes all important the expected profits under the two
organizational structures converge.
27
For the Delegation Principle see, for instance, Milgrom and Roberts (1992).

21
 δ
1+δ
1
1.0

0.9

0.8

0.7 Centralization
0.6

0.5

0.4

0.3
Decentralization
0.2
δ /(1 + δ )
0.1

00.0
0.5
1/2
0.6
17/28
0.7 0.8 0.9 1.0
1 λ

Figure 5: Organizational Performance

To understand these results, it is useful to decompose the effect of a change in δ on the expected
profits of the two organizational structures:
µ µ ¶ µ ¶ ¶
dΠl ∂ALl ∂CLl ∂ALl ∂CLl dSl
= − CLl + + 2δ + + 2δ for l = C, D. (17)
dδ ∂δ ∂δ ∂Sl ∂Sl dδ

The first term denotes the direct effect of an increase in δ on the expected profits, holding constant
decision making and communication: an increase in δ puts more weight on the coordination loss
and thus reduces the expected profits under both organizational structures. Since, from Lemma 2,
CLC ≤ CLD , this effect makes Centralization more attractive relative to Decentralization.
The second term reflects how an increase in δ affects decision making holding constant com-
munication and the weight on the coordination loss in the profit function. Under Centralization
this effect is equal to zero while under Decentralization it is negative for small values of δ and
positive when δ is sufficiently big.28 Intuitively, whereas for δ = 0 division managers are unbiased,
for δ > 0 they put too little weight on coordination. For δ large enough, however, more interdepen-
dence aligns the incentives of the decision makers. Indeed, recall from Section 4 that in the limit
as coordination becomes all important, decision making under Centralization and Decentralization
converge.
28
The zero effect under Centralization is due to the Envelope Theorem: an increase in δ induces the HQ Manager
to change her decision making so as to reduce the coordination loss CLC and increase adaptation loss ALC ; the
change in decision making, however, does not have a first order effect on the expected profits.

22
 Finally, the last term represents the impact of an increase in δ on communication, holding
constant decision making and the weight on the coordination loss in the profit function. Under
Centralization this effect is always negative: an increase in δ leads to worse communication which
reduces the expected profits. In contrast, under Decentralization this effect is always positive
since an increase in δ actually improves communication and thereby increases the expected profits.
In a model with non-strategic communication the first two terms on the RHS of (17) would still
be present but the third term would not. We show below that the fact that the quality of
communication is endogenous in our model, and therefore the third term is in general not zero, is
key for the results in Proposition 5.

δ small Consider first how the different effects play out when δ is small. Intuitively, even when
the need for coordination is very small communication is noisy which affects the ability of the
HQ Manager to adapt the decisions to the local conditions and the ability of division managers
to coordinate them. Since for small δ coordination is much less important than adaptation, the
centralized structure suffers more from imperfect communication than the decentralized one. As
a result, Decentralization dominates Centralization.
To see this intuition more formally, note that for δ = 0, d1 = θ1 and d2 = θ2 under both
organizational structures. The coordination loss, and thus the first term on the RHS of (17), is
£ ¤
then the same under both structures: CLC = CLD = E (θ1 − θ2 )2 . The second term on the
RHS of (17), which captures the effect of changes in the decision making on the expected profits,
is also the same. This is so since, at δ = 0, decision making is efficient under both organizational
structures and as a result small changes in the decision making do not have a first order effect on
the expected profits. The differences in the expected profits for small δ are therefore determined
by how a change in the need for coordination affects communication, i.e. the third term on the
RHS of (17). In particular, for δ = 0
µ ¶ µ ¶
dΠD dΠC ∂ALC ∂CLC dSC ∂ALD ∂CLD dSD
− = + 2δ − + 2δ
dδ dδ ∂SC ∂SC dδ ∂SD ∂SD dδ
∂ALC dSC
= > 0.
∂SC dδ
Under Centralization an increase in δ results in a deterioration in communication which limits the
HQ Manager’s ability to adapt the decisions to the local conditions. There is no such adaptation loss
under Decentralization since division managers do not rely on communication for their adaptation
decisions. Changes in the quality of communication also affect the ability of the decision makers
to coordinate the decisions. This effect is given by the terms 2δ∂CLl /∂Sl , l = C, D, in above

23
expression which is zero at δ = 0. Thus, for small δ Decentralization dominates Centralization
since noisy communication has a more detrimental effect on the HQ Manager’s ability to adapt
than on the division managers’ abilities to coordinate.

δ large The quality of communication is also key for understanding the relative performance of
the two organizational structures when coordination is very important. To see this, consider first
what happens in the limit in which δ → ∞. We know from Section 4 that in this case the decision
making under the two structures converge. Essentially, when coordination becomes sufficiently
important the division managers recognize their interdependence and set both decisions equal to
the average state, just like the HQ Manager. Since we know from Section 5 that the quality
of communication also converges in the limit, it is intuitive that limδ→∞ ΠC = limδ→∞ ΠD , as
shown in Part (iii.) of the above proposition. If the quality of communication did not converge,
then Centralization would always dominate Decentralization in the limit. In particular, fixing the
quality of communication at some level SC and SD and taking the limit of the expected profits as
δ goes to infinity yields:
1 ¡ ¢ 1 ¡ ¢
lim ΠD |SD = − (1 + SD ) σ 21 + σ 22 < − (1 + SC ) σ 21 + σ 22 = lim ΠC |SC , (18)
δ→∞ 2 2 δ→∞

where the inequality follows from SD > SC for δ ∈ (0, ∞) and λ ∈ (1/2, 1]. Moreover, since
the expected profits are continuous in δ, (18) implies that for sufficiently high but finite values of
δ Centralization would always strictly dominate Decentralization if the quality of communication
were held constant. Thus, both the fact that Decentralization can be optimal for any finite δ > 0
and the fact that the two structures perform equally well when δ → ∞ are driven by the endogenous
quality of communication. In other words, if communication were non-strategic, so that the third
term on the RHS of (17) were zero, coordination would require centralization.
To shed more light on why Decentralization can be optimal for any finite δ > 0, it is useful to
decompose the effect of a change in λ on the expected profits of the two organizational structures
as
dΠl ∂Πl ∂Πl dSl
= + for l = C, D.
dλ ∂λ ∂Sl dλ
The first term relates to the impact of λ on decision making, keeping the communication constant,
and the second term reflects the impact of a change in λ on the quality of communication. Recall
that at λ = 1/2 both organizational structures achieve the first best expected profits. As a result,
the first term is zero under both organizational structures when λ = 1/2. For small λ, therefore,
the preferred organizational structure will be the structure which suffers least from the deterioration
in the quality communication.

24
 δ
1+ δ

Figure 6: The Relative Value of Decentralization: ΠC − ΠD

Under both organizational structures even a very small own-division bias has a first order
impact on communication quality, that is for λ = 1/2, dSl /dλ > 0 for l = C, D. A deterioration
in communication quality affects the ability of the HQ Manager to adapt to local information
and it affects the ability of the division managers to coordinate their decisions. The impact of an
information loss on the HQ Manager’s ability to adapt, however, is bigger than its impact on the
division managers’ ability to coordinate. Indeed, one can verify that, for λ = 1/2,

∂ΠC dSC ∂ΠD dSD

< < 0.
∂SC dλ ∂SD dλ
It follows that for any finite δ > 0, Decentralization strictly dominates Centralization when the
own-division bias λ > 1/2 is sufficiently small. The fact that the Delegation Principle is satisfied
in our model, therefore, is also due to the endogenous quality of communication.

6.1 The Relative Value of Centralization

So far we have focused on the information cost of centralization and have abstracted from any
resource costs. It is evident, however, that setting up a headquarters and hiring a manager to run
it involves additional resource costs, at the very least the opportunity cost of the additional manager.
In this sub-section we briefly digress from our main analysis to investigate the comparison between
Centralization and Decentralization when setting up a headquarters involves such a resource cost.

25
 δ
1+δ
1
1.0

0.9

0.8

0.7

0.6
Centralization
0.5
Decentralization
0.4

0.3

0.2

0.1

0
0.0
0.5
1/2 0.6 0.7 0.8 0.9 1.0
1 λ

Figure 7: Organizational Performance with Resource Costs

¡ ¢
For this purpose, consider first Figure 6 which plots ΠC − ΠD for σ 21 + σ 22 = 1.29 Confirming
our previous results, it can be seen that ΠC − ΠD < 0 if either the own-division bias or the need
for coordination are small and that ΠC − ΠD ≥ 0 otherwise. Suppose now that Centralization
involves a resource cost, namely a fixed wage w > 0 for the HQ Manager. The expected profits
under Centralization are then given by ΠC = ΠC − w. It is evident that once we allow for a
resource cost Decentralization will still dominate Centralization when either λ or δ are sufficiently
small. However, now Decentralization also dominates Centralization for very high values of δ.
This is the case since, as δ becomes very large, ΠC − ΠD converges to zero which, of course, implies
that ΠC − ΠD becomes negative. To see this graphically, consider again Figure 6 and note that
ΠC − ΠD can be obtained by simply shifting ΠC − ΠD down by w.
Consider next Figure 7 which compares the performance of the two organizational structures
for w = 0.01. The figure illustrates two key features of the model with resource costs. First,
Decentralization dominates Centralization when the need for coordination is either very small or
very big and, provided that the own-division bias is sufficiently big, Centralization dominates for
intermediate values of δ. Second, the effect of an increase in δ on the dominant organizational
structure is ambiguous: an increase in δ can make it optimal to switch from Decentralization to
Centralization or vice versa.
29
   
Since both ΠC and ΠD are proportional to σ 21 + σ22 setting σ21 + σ22 = 1 is without loss of generality.

26
6.2 Information Aggregation

Headquarters are often decried as costly and badly informed bureaucracies that needlessly inter-
fere with the decisions of better informed division managers. This sentiment reflects the natural
information advantage of division managers who learn relevant information directly through their
involvement in the day-to-day operations of an organization while managers at headquarters rely
on the communication of this information by the division managers. However, since operating
managers are more willing to share information with headquarters than with each other, managers
at headquarters enjoy a counteracting communication advantage. As a result, a headquarters man-
ager who does not observe any information directly may still be a better aggregator of information
than a division manager who does observe some information himself but also relies on communi-
cation to obtain other information. In this section we investigate whether this can indeed be the
case. So far we allowed for differences in the variances of the local conditions σ 21 and σ 22 . For
expositional convenience we now restrict attention to the case in which the two division managers
are equally informed, i.e. σ 21 = σ 22 , and define σ 2 ≡ σ 21 = σ 22 .
There are many ways in which one could measure how well a manager aggregates information.
A natural measure, and the one we focus on, is the average accuracy of a manager’s posterior beliefs
about each state as given by the average residual variance of θ1 and θ2 . We know from the analysis
above that under Centralization the average residual variance of the HQ Manager is given by SC σ 2
while under Decentralization the average residual variance of each division manager is given by
SD σ 2 /2. The information advantage of the division managers is reflected by the residual variance
of their own state being zero while the HQ Manager’s communication advantage is reflected by
the fact that SC ≤ SD . The next proposition shows under what conditions the communication
advantage dominates the information advantage.

PROPOSITION 6 (Information Aggregation). Suppose that σ 21 = σ 22 . Then SC ≤ SD /2 for all

δ ∈ [0, δ 1 ] and SC ≥ SD /2 for all δ ∈ [δ 1 , ∞) , where

λ ³ p ´
δ1 ≡ 2 − λ + 20λ + λ2 + 1 > 0.
8λ − 1

Thus, the HQ Manager is a better aggregator if and only if coordination is not very important.
In light of our previous discussion the intuition for this result is straightforward: for small δ
the HQ Manager’s communication advantage is significant and dominates the division managers’
information advantage; as δ increases, however, the communication advantage continuously shrinks
and, eventually, is dominated by the information advantage.

27
7 Asymmetric Organizations

A key feature of our model is the symmetry of the organization: the two divisions are of equal size,
they have the same need for coordination and the two decisions are made simultaneously. In this
section we relax the symmetry assumptions and investigate how asymmetries between the divisions
affect the relative performance of centralized and decentralized organizations. We show that such
asymmetries tend to favor centralization when coordination is very important. However, provided
that the differences between the divisions are not too big, it is still the case that Decentralization
can be optimal even if the need for coordination is arbitrarily high. Our key results therefore
continue to hold, albeit in a weaker form. To streamline the exposition, and in contrast to our
previous analysis, we now assume that both division managers have the same amount of private
information, in the sense that σ 21 = σ 22 . Also, we relegate the formal analysis to the appendix (see
Appendices B, C and D).

7.1 Sequential Decision Making

The first asymmetry we consider concerns the timing of decision making under Decentralization.
In many settings one of the division managers may have a first mover advantage, that is he may be
able to make his decision before the other division manager can do so. To explore this possibility,
we now consider a version of our model in which the decision making under Decentralization
takes place sequentially. In particular, suppose that Manager 1 is the Leader and Manager 2 the
Follower. After both division managers have observed their local conditions, the Follower sends a
single message to the Leader. Next the Leader makes his decision, the Follower observes it and
then makes his own.
At first it would seem that sequential decision making facilitates coordination since it eliminates
the Follower’s uncertainty about the Leader’s decision. This would suggest that allowing for
sequential decision making further strengthens the result that Decentralization can be optimal
even when coordination is very important. There are, however, two countervailing effects. First,
under sequential decision making the Leader is more uncertain about the Follower’s decision than
he is under simultaneous decision making. This is the case since the Follower communicates less
information to the Leader than he does to a peer. Second, when decision making takes place
sequentially the Leader adapts his decision more closely to his local conditions since he knows that
the Follower will be forced to adjust his decision accordingly. As a result, decisions can be less
coordinated under sequential than under simultaneous decision making. The next proposition
investigates how these different effects play out when either coordination is very important or the

28
own-division bias is very small.

PROPOSITION 7 (Sequential Decision Making).

i. For any λ ∈ (1/2, 1] Centralization strictly dominates Decentralization with sequential
decision making when coordination is sufficiently important.
ii. For any δ ∈ (0, ∞) Decentralization with sequential decision making strictly dominates
Centralization when the own-division bias λ > 1/2 is sufficiently small.
Part (i.) highlights a key difference between the model with sequential decision making and the
symmetric model while Part (ii.) highlights a key similarity. In particular, Part (i.) shows that
when decision making takes place sequentially and coordination is very important, then efficient
coordination does require centralization. This is the case since even in the limit in which δ → ∞
decision making under Decentralization is biased and headquarters’ communication advantage does
not vanish. In contrast, when decision making takes place simultaneously then, in the limit, decision
making and communication are as efficient under Decentralization as under Centralization and the
two structures perform equally well.
However, it is still the case that, no matter how important coordination, Decentralization strictly
dominates Centralization if the own-division bias λ > 1/2 is sufficiently small, as shown in Part (ii.)
of the proposition. The reason is the same as when decision making takes place simultaneously:
for λ close to 1/2 the information advantage of the decentralized structure dominates the commu-
nication advantage of the centralized one. The central insight of the symmetric model therefore
continues to hold in this extension, albeit in a weaker form. In other words, coordination does
not necessarily require centralization, even when decisions are made sequentially.

7.2 Different Needs for Coordination

In many organizations divisions differ in how much their profits depend on coordination. To allow
for this possibility we now extend the model considered in Section 3 by assuming that the weight
that Division 1 puts on coordination is given by δ 1 ∈ [0, ∞) while that of Division 2 is given by
δ 2 ∈ [0, ∞). As in the previous sub-section, we again focus on the relative performance of the
two organizational structures when either coordination is very important or the own-division bias
is very small:30

PROPOSITION 8 (Different Needs for Coordination).

30
For a full analysis of this model when λ = 1 see Rantakari (2006).

29
 i. For any λ ∈ (1/2, 1] and δ j ∈ [0, ∞), j = 1, 2, Centralization strictly dominates Decen-
tralization when coordination is sufficiently important for Division k 6= j.
ii. For any δ 1 , δ 2 ∈ (0, ∞) Decentralization strictly dominates Centralization when the own-
division bias λ > 1/2 is sufficiently small.
Part (i.) again highlights a key difference with the symmetric model while Part (ii.) highlights
a key similarity. In particular, Part (i.) shows that when coordination becomes very important
for one of the divisions but not the other, then efficient coordination does require centralization.
In spite of this fact, however, the central result of the symmetric model continues to hold, as
shown in Part (ii.): no matter how important coordination is for each division, Decentralization
dominates Centralization when the own-division bias λ > 1/2 is sufficiently small. Thus, even
when the divisions have different needs for coordination, coordination does not necessarily require
centralization.

7.3 Different Division Sizes

Finally, we allow for differences in the size of the two divisions. In particular, we now extend
the model considered in Section 3 by assuming that the profits of Divisions 1 and 2 are given by
2απ 1 and 2 (1 − α) π 2 respectively, where π 1 and π 2 are defined in (1) and (2) and α ∈ [1/2, 1)
is a parameter that measures the relative size of the two divisions. Once again we focus on the
performance of the two organizational structures when either the need for coordination is very big
or the own-division bias is very small:

PROPOSITION 9 (Different Division Sizes).

i. For any λ > 1/2 and α > 1/2 Centralization strictly dominates Decentralization when
coordination is sufficiently important.
ii. For any δ ∈ (0, ∞) Decentralization strictly dominates Centralization when the own-div-
ision bias λ > 1/2 and the difference in the division sizes α > 1/2 are sufficiently small.
Thus, as in the two previous extensions, Centralization strictly dominates Decentralization for
any λ > 1/2 when coordination is sufficiently important. Moreover, provided that the difference
in the division sizes is sufficiently small, it is again the case that for any finite δ, Decentralization
strictly dominates Centralization if the own-division bias λ > 1/2 is sufficiently small. If, however,
the difference in the division sizes is ‘too big,’ then Centralization can be optimal even if the
incentives of the division managers are extremely closely aligned, i.e. if λ > 1/2 is arbitrarily close
to 1/2. The extent to which coordination requires centralization therefore depends on the relative
size of the divisions. Essentially, the bigger the size difference between the divisions, the more

30
likely it is that coordination requires centralization.

8 Conclusions

When does coordination require centralization? In this paper we addressed this question in a model
in which information about the costs of coordination is dispersed among division managers who
communicate strategically to promote their own divisions at the expense of the overall organization.
We showed that vertical communication is more efficient than horizontal communication: division
managers share more information with an unbiased headquarters than they do with each other.
However, headquarters’ communication advantage diminishes, and eventually vanishes, as coordi-
nation becomes more important. As a result, decentralization can be optimal even if coordination
is very important. In particular, this is so if the incentives of the division managers are sufficiently
aligned and the divisions are not too different from each other. Thus, only in organizations in
which divisions are quite heterogenous, for instance in their relative size, or in which implicit and
explicit incentives strongly bias division managers towards their own divisions, does coordination
require centralization.
The analysis of our model is surprisingly simple and, arguably, more tractable than the leading
example in Crawford and Sobel (1982), the traditional workhorse for cheap talk applications. This
is so since in our setting the efficient communication equilibrium is one in which the number of
intervals goes to infinity. As a result, one can avoid the integer problem that is associated with
the finite interval equilibria in the standard model. We hope that our framework will prove useful
in approaching various applied problems ranging from organizational design, the theory of the firm
to fiscal federalism.
Our central result — that decentralization can be optimal even if coordination is very important
— is reminiscent of Hayek (1945) and the many economists since who have invoked the presence of
‘local knowledge’ as a reason to decentralize decision making in organizations. In the management
literature, this view has become known as the Delegation Principle.31 Our insight differs from the
standard Delegation Principle rationale on two important dimensions.
First, the standard argument posits that efficient decision making requires the delegation of
decision rights to those managers who possess the relevant information. The alternative — commu-
nicating the relevant information to those who possess the decision rights — is discarded on the basis
of physical communication constraints (Jensen and Meckling 1992). Is the local knowledge argu-
ment still relevant in a world in which technological advances have slashed communication costs?
31
See, for instance, Milgrom and Roberts (1992).

31
Our analysis suggests that it is: even in the absence of any physical communication costs, local
knowledge remains a powerful force for decentralization. The same factors that make decentraliza-
tion of decision rights unattractive — biased incentives of the local managers — are even more harmful
for the transfer of information. As long as division managers are not too biased, the distortion of
information in a centralized organization outweighs the loss of control under decentralization.
Second, the standard rationale for the Delegation Principle presumes that decentralization
eliminates the need for communication. If, however, decision relevant information is dispersed
among multiple managers — as is likely to be the case when decisions need to be coordinated
— then efficient decision making always requires communication. Our analysis has highlighted
that while local knowledge gives division managers an information advantage, headquarters has a
communication advantage. As a result, headquarters can actually be a more efficient aggregator of
dispersed information. Despite the need for information aggregation, we show that decentralization
is often preferred, even if coordination is very important. The reason is that as coordination and
information aggregation become more important, division managers communicate more efficiently
with each other and headquarters’ communication advantage diminishes and eventually disappears.
Beyond its implication for organizational design, it is tempting to interpret our theory as one of
horizontal firm boundaries. In order to have a compelling theory of horizontal integration, however,
it would be necessary to combine our approach with the possibility of some bargaining between
division managers or between division managers and headquarters. It would further be important
to endogenize managerial incentives, as managers in independent firms may have different objectives
than those belonging to the same organization. These issues, and others, await future research.

32
9 Appendix A

We first define the random variable mi as the posterior expectation of the state θi by the receiver
of message mi . The following lemma will be used throughout the appendix.

LEMMA A1. For any communication equilibrium considered in Proposition 1 Eθ2 [m1 m2 ] =
Eθ2 [θ1 m2 ] = Eθ2 [θ2 ] = Eθ2 [m2 ] = Eθ2 [θ1 θ2 ] = 0 and Eθ1 [m1 m2 ] = Eθ1 [θ2 m1 ] = Eθ1 [θ1 ] =
Eθ1 [m1 ] = Eθ2 [θ1 θ2 ] = 0.

Proof: All equalities follow from independence of θ1 and θ2 and that in equilibrium Eθi [mi ] =
Eθi (θi ) = 0. ¥

Proof of Proposition 1: We first note that in any Perfect Bayesian Equilibrium of the com-
munication game, optimal decisions given beliefs satisfy (3) and (4) for the case of Centralization
and (8) and (9) for the case of Decentralization. Now we will establish that communication rules
in equilibrium are interval equilibria.
For the case of Centralization let μ2 (·) be any communication rule for Manager 2. The expected
utility of Manager 1 if the HQ Manager holds a posterior expectation ν 1 of θ1 is given by
µ ³ ´2 ³ ´2 ³ ´2 ¶
bC bC
Eθ2 (U1 | θ1 , ν) = −Eθ2 λ d1 − θ1 + (1 − λ) d2 − θ2 + δ d1 − d2bC bC
, (19)

with

dbC
1 ≡ γ C ν 1 + (1 − γ C ) E (θ2 | μ2 (·)) (20)
dbC
2 ≡ (1 − γ C ) ν 1 + γ C E (θ2 | μ2 (·)) . (21)

∂2 ∂2
It is readily seen that ∂θ1 ∂ν 1 Eθ2 (U1 | θ1 , ν 1 ) > 0 and E
∂ 2 θ1 θ 2
(U1 | θ1 , ν 1 ) < 0. This implies
that for any two different posterior expectations of the HQ Manager, say ν 1 < ν 1 , there is at most
one type of Manager 1 that is indifferent between both. Now suppose that contrary to the assertion
¡ ¢ ¡ ¢
of interval equilibria there are two states θ11 < θ21 such that Eθ2 U1 | θ11 , ν 1 ≥ Eθ2 U1 | θ11 , ν 1 and
¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢
Eθ2 U1 | θ21 , ν 1 > Eθ2 U1 | θ21 , ν 1 . But then Eθ2 U1 | θ21 , ν 1 −Eθ2 U1 | θ21 , ν 1 < Eθ2 U1 | θ11 , ν 1 −
¡ ¢ 2
Eθ2 U1 | θ11 , ν 1 which violates ∂θ∂1 ∂ν 1 Eθ2 (U1 | θ1 , ν 1 ) > 0 . The same argument can be applied to
Manager 2 for any reporting strategy μ1 (·) of Manager 1. Therefore all equilibria of the communi-
cation game under Centralization must be interval equilibria.
For the case of Decentralization let μ1 (·) and μ2 (·) be communication rules of Manager 1 and
Manager 2, respectively. Sequential rationality implies that in equilibrium decision rules must
conform to (8) and (9). If ν 1 denotes the expectation of the posterior of θ1 that Manager 2 holds
∂2
then ∂θ1 ∂ν 1 Eθ2 (U1 | θ1 , ν 1 ) > 0 and the proof follows as in the preceding paragraph.

33
 We now characterize all finite equilibria of the communication game, i.e. equilibria that induce a
finite number of different decisions. For this purpose for Manager j let aj be a partition of [−sj , sj ],
any message mj ∈ (aj,i−1 , aj,i ) be denoted by mj,i and mj,i be the receiver’s posterior belief of the
expected value of θj after receiving message mj,i .
a. Centralization: In state a1,i Manager 1 must be indifferent between sending a message that
induces a posterior m1,i and a posterior m1,i+1 so that Eθ2 (U1 | a1,i , m1,i ) − Eθ2 (U1 | a1,i , m1,i+1 ) =
0. Using Lemma A1 on (19) we have that

Eθ2 (U1 | a1,i , m1,i ) − Eθ2 (U1 | a1,i , m1,i+1 )

= λ (γ C m1,i+1 − a1,i )2 + (1 − λ) (1 − γ C )2 m21,i+1 + δ (2γ C − 1)2 m21,i+1
³ ´
− λ (γ C m1,i − a1,i )2 + (1 − λ) (1 − γ C )2 m21,i + δ (2γ C − 1)2 m21,i .

Substituting m1,i = (a1,i−1 + a1,i ) /2 we have that Eθ2 (U1 | a1,i , m1,i ) − Eθ2 (U1 | a1,i , m1,i+1 ) = 0 if
and only if a1,i = (a1,i−1 + a1,i+1 ) / (2 + 4bC ), where bC is defined in the proposition. Rearranging
this expression, we get
a1,i+1 − a1,i = a1,i − a1,i−1 + 4bC a1,i . (22)

Let the total number of elements of a1 be N . Equilibrium with an even number of elements
would correspond to the case N = 2N1 , and N = 2N1 + 1 when the equilibrium has an odd number
of elements. Using the boundary conditions a1,0 = −s1 and a1,N = s1 to solve the difference
equation (22) gives
s1 ¡ ¢
a1,i = ¡ N N
¢ xiC (1 + yC
N i
) − yC (1 + xN
C ) for 0 ≤ i ≤ N, (23)
xC − yC
q
where the roots xC and yC are given by xC = (1 + 2bC ) + (1 + 2bC )2 − 1 and yC = (1 + 2bC ) −
q
(1 + 2bC )2 − 1 and satisfy xC yC = 1, with xC > 1. It is readily seen that for each k ≤ N,
a1,k = a1,N−k , i.e. the intervals are symmetrically distributed around zero. When N = 2N1 , i.e.
the partition has an even number of elements, then a1, N = 0. In this case we can compactly write
2
(23) as
s1 ¡ ¢
a1,i = ³ ´ xiC − yC
i
) for 0 ≤ i ≤ N1
xN1
C − yC
N1

For an equilibrium with an odd number of elements N = 2N1 + 1 there is a symmetric interval
around zero where the HQ Manager’s expected posterior of θ1 is zero. In this case we can compactly
write (23) as
s1 ³ ´
a1,i = ³ ´ xiC (xN
C
1
+ yC
N1 +1
) − y i
C C(xN1 +1
+ yC
N1
)) for 1 ≤ i ≤ N1 .
2N1 +1 2N1 +1
xC − yC

34
The analysis for Manager 2 is analogous.
b. Decentralization: If Manager 1 observes state θ1 and sends a message m1,i that induces a
posterior belief m1,i in Manager 2 his expected utility is given by
³ ¡ ¢2 ¡ ¢2 ¡ ¢ ´
D 2
Eθ2 (U1 | θ1 , m1,i ) = −Eθ2 λ dD
1 − θ1 + (1 − λ) dD
2 − θ2 + δ dD
1 − d2 , (24)

where dD D
1 and d2 are given by (8) and (9). It must again be the case that Eθ2 (U1 | a1,i , m1,i ) −
Eθ2 (U1 | a1,i , m1,i+1 ) = 0. Making use of Lemma A1 on (24) we have that

Eθ2 (U1 | a1,i , m1,i ) − Eθ2 (U1 | a1,i , m1,i+1 )

à !
2 (m1,i+1 δ − (λ + 2δ) a1,i )
2 δ 2 m21,i+1 λ2 (m1,i+1 δ − (λ + 2δ) a1,i )2
= λδ + (1 − λ) +δ
(λ + δ)2 (λ + 2δ)2 (λ + 2δ)2 (λ + δ)2 (λ + 2δ)2
à !
2 (m1,i δ − (λ + 2δ) a1,i )
2 δ 2 m21,i λ2 (m1,i δ − (λ + 2δ) a1,i )2
− λδ + (1 − λ) +δ
(λ + δ)2 (λ + 2δ)2 (λ + 2δ)2 (λ + δ)2 (λ + 2δ)2

Substituting m1,i = (a1,i−1 + a1,i ) /2 and m1,i+1 = (a1,i + a1,i+1 ) /2 we have that Eθ2 (U1 | a1,i , m1,i )−
Eθ2 (U1 | a1,i , m1,i+1 ) = 0 if and only if a1,i = (a1,i−1 + a1,i+1 ) / (2 + 4bD ), where bD is defined in
the proposition. Rearranging this expression, we get

a1,i+1 − a1,i = a1,i − a1,i−1 + 4bD a1,i . (25)

Using the boundary conditions a1,0 = −s1 and a1,N = s1 to solve the difference equation(25) gives
s1 ¡ i N i N
¢
a1,i = ¡ N N
¢ xD (1 + yD ) − y D (1 + xD ) for j = 1, 2 and 0 ≤ i ≤ N, (26)
xD − yD
q
where the roots xD and yD are given by xD = (1 + 2bD ) + (1 + 2bD )2 − 1 and yD = (1 + 2bD ) −
q
(1 + 2bD )2 − 1 and satisfy xD yD = 1, with xD > 1. It is readily seen that for each k ≤ N,
a1,k = a1,N−k , i.e. the intervals are symmetrically distributed around zero. When N = 2N1 , i.e.
the partition has an even number of elements, then a1, N = 0. In this case we can compactly write
2
(26) as
s1 ¡ ¢
a1,i = ³ ´ xiD − yD
i
) for 0 ≤ i ≤ N1
xN1
D − yD
N1

For an equilibrium with an odd number of elements N = 2N1 + 1 there is a symmetric interval
around zero where the HQ Manager’s expected posterior is zero. In this case we can compactly
write (26) as
s1 ³ ´
a1,i = ³ ´ xiD (xN
D
1
+ yD
N1 +1
) − y i
D D(xN1 +1
+ yD
N1
)) for 1 ≤ i ≤ N1 .
2N1 +1 2N1 +1
xD − yD

35
The analysis for Manager 2 is analogous. ¥

For the proof of Proposition 2 we will make use of the following lemma:
LEMMA A2. i. E(mj θj ) = E(m2j ) for j = 1, 2. ii. E(m2j ) is strictly increasing in the num-
ber of intervals Nj for j = 1, 2. iii. Under Centralization lim E(m2j ) = (1 − SC ) σ 2j and under
Nj →∞
Decentralization lim E(m2j ) = (1 − SD ) σ 2j , where SC and SD are defined in (13).
Nj →∞

Proof: i. Given the equilibrium reporting strategies we have that mj = E [θj |θj ∈ (aj,i−1 , aj,i ) ] for
j = 1, 2, which implies

E(mj θj ) = E [E [mj θj |θj ∈ (aj,i−1 , aj,i ) ]] = E [mj E [θj |θj ∈ (aj,i−1 , aj,i ) ]] = E(m2j ).

ii. Using (23) we can compute E(m2j ) as follows

Z µ ¶
1 X aj,i aj,i + aj,i−1 2 1 X
2
E(mj ) = dθj = (aj,i − aj,i−1 ) (aj,i + aj,i−1 )2
2sj aj,i−1 2 8s1
Nj Nj

s2j X ³ ´
N 3 3(i−1)
³ ´
N 3 3(i−1)
= ³ ´3 { 1 + yC j xC (xC + 1)2 (xC − 1) + 1 + xC j yC (yC + 1)2 (1 − yC )
Nj Nj
8 xC − yC Nj
³ ´³ ´ ³ ´ ³ ´
N N 2 i−1 2 N 2 N
1 + yC j 1 + xC j yC (yC − 1)(xC + 1) − 1 + yC j 1 + xC j xi−1 2
C (xC − 1)(yC + 1)}.

Performing the summation in the above expression and using the fact that xC yC = 1 we get
after some lengthy calculations that

⎡
³ ´ ³ ´2 ⎤
3Nj 2 Nj
s2j⎢ x C − 1 (xC − 1) xC + 1 (yC + 1)(xC + 1) ⎥
E(m2j ) = ⎣³ ´3 ¡ ¢ − ³ ´2 ⎦= (27)
4 N N N N
xC j − 1 x2C + xC + 1 xC j xC j − yC j
⎡ ³ ´ ⎤
3Nj 2 N
2
sj ⎢ xC − 1 (xC − 1) xC (xC + 1)2 ⎥
j

= ⎣³ ´3 ¡ ¢ − ³ ´2 ⎦ .
4 Nj 2 Nj
xC − 1 xC + xC + 1 xC xC − 1

To see that E(m2j ) is strictly increasing in Nj first define

p3 − 1 (xC − 1)2 p (xC + 1)2

f (p) = ¡ ¢ −
(p − 1)3 x2C + xC + 1 (p − 1)2 xC
and note that
p+1 (xC − 1)2 p + 1 (xC + 1)2
f 0 (p) = −3 ¡ ¢ + =
(p − 1)3 x2C + xC + 1 (p − 1)3 xC
p + 1 10x2C + x4C + 1
= ¡ ¢.
(p − 1)3 xC x2C + xC + 1

36
 s2j Nj
Therefore f 0 (p) > 0 for p > 1. Since E(m2j ) = 4 f (xC ) this establishes that E(m2j ) is strictly
increasing in Nj .
iii. Taking limits to each term on the RHS of (27) we obtain
³ ´
3N
xC j − 1 (xC − 1)2 xC − 1
lim ³ ´3 ¡ ¢ = 2
Nj →∞ N xC + xC + 1
xC j − 1 x2C + xC + 1
N
xC j (xC + 1)2
lim ³ ´2 = 0.
Nj →∞ N
xC xC j − 1

Therefore
s21 (xC + 1)2 1 + bC
lim E(m2j ) = ¡ 2 ¢ = s21 = (1 − SC ) σ 2j .
Nj →∞ 4 xC + xC + 1 3 + 4bC
The analysis for Decentralization is analogous (one just needs to replace C with D in the above
expressions). ¥

Proof of Proposition 2: First we establish that the limit of strategy profiles and beliefs (μ1 (·),
μ2 (·) , d1 (·), d2 (·), g1 (·), g2 (·)) as N1 , N2 → ∞ denoted by (μ∞ ∞ ∞ ∞ ∞ ∞
1 (·), μ2 (·) , d1 (·), d2 (·), g1 (·), g2 (·))
is indeed a Perfect Bayesian Equilibrium of the communication game, both under Centralization and
under Decentralization. To this end we need only verify that the reporting strategies of Manager
1 and Manager 2 are a best response to (d∞ ∞
1 (·), d2 (·)). Now suppose there is a θ j that induces an
expected posterior m1 in the decision maker and has a profitable deviation by inducing a different
expected posterior m2 associated with state θ0j . But then there exists a finite Nj such that θj has
a profitable deviation by inducing the same posterior that θ0j contradicting the fact that strategy
profiles and beliefs (μ1 (·), μ2 (·) , d1 (·), d2 (·), g1 (·), g2 (·)) constitute an equilibrium for all finite Nj .
Thus θj cannot profitably deviate and therefore the reporting strategies (μ∞ ∞
1 (·), μ2 (·)) are a best
response to (d∞ ∞
1 (·), d2 (·)). The fact that the expected profit under the equilibrium with an infinite
number of elements for Manager 1 and Manager 2 coincides with the limit of the expected profit
obtains by applying the Lebesgue Dominated Convergence Theorem to the expression of Πl with
l = {C, D} .
Finally, we show that the equilibrium (μ∞ ∞ ∞ ∞ ∞ ∞
1 (·), μ2 (·) , d1 (·), d2 (·), g1 (·), g2 (·)) yields higher
total expected profits than all finite equilibria .
a. Centralization: The expected profits under Centralization are given by
³¡ ¢2 ¡ C ¢2 ¡ C ¢ ´
C 2
ΠC = −E dC
1 − θ 1 + d2 − θ 2 + 2δ d1 − d2 (28)

37
Using (3), (4) and Lemmas A1 and A2-i. we have that for given N1 , N2
³¡ ¢2 ´ ¡ ¢ ¡ ¢
E dC = σ 21 − γ C (2 − γ C ) Eθ1 m21 + (1 − γ C )2 Eθ2 m22
1 − θ1 (29)
³¡ ¢2 ´ ¡ ¢ ¡ ¢
E dC2 − θ2 = σ 22 + (1 − γ C )2 Eθ1 m21 − γ C (2 − γ C ) Eθ2 m22
³¡ ¢ ´ ¡ ¡ ¢ ¡ ¢¢
C 2
C
E d1 − d2 = (2γ C − 1)2 Eθ1 m21 + Eθ2 m22
³ ´
The rate at which total profits change with the variance of the messages Eθj m2j is given by

∂ΠC
³ ´ = (1 + 2δ) (2γ C − 1)
∂Eθj m2j
³ ´
1 ∂ΠC
Since γ C > 2 we have that ∂Eθj (m2j )
> 0. From Lemma A.2-ii we have that Eθj m2j increases
with Nj and therefore expected profits ΠC increase as the number of elements Nj in the partition
of Manager j increases.
b. Decentralization: The expected profits under Decentralization are given by
³¡ ¢2 ¡ D ¢2 ¡ D ¢ ´
D 2
ΠD = −E dD
1 − θ 1 + d2 − θ 2 + 2δ d1 − d2 . (30)

Using (8), (9) and Lemmas A1 and A2-i. we have that for given N1 , N2
³¡ ¢2 ´ δ2 2λ + 3δ ¡ 2¢ δ2 ¡ 2¢
E dD
1 − θ 1 = σ
2 1
2
− δ 3
2 2 Eθ 1 m 1 + 2 Eθ2 m2 (31)
(λ + δ) (λ + δ) (λ + 2δ) (λ + 2δ)
³¡ ¢2 ´ δ2 δ2 ¡ 2¢ 2λ + 3δ ¡ 2¢
E dD 2 − θ 2 = σ
2 2
2
+ 2 Eθ1 m1 − δ
3
2 2 Eθ2 m2
(λ + δ) (λ + 2δ) (λ + δ) (λ + 2δ)
³¡ ¢ ´ λ2 ¡ 2 ¢ 2λ + 3δ ¡ ¡ 2¢ ¡ 2 ¢¢
D 2
E dD 1 − d2 = 2
2 σ1 + σ2 − λ δ
2
2 2 Eθ1 m1 + Eθ2 m2
(λ + δ) (λ + δ) (λ + 2δ)
³ ´
The rate at which total profits change with the variance of the messages Eθj m2j is
à µ ¶2 !
∂ΠD λ2 δ 2 δ
³ ´= 6δ − 1 + 4λ + 2
∂Eθj m2j (λ + δ)2 (λ + 2δ)2 λ

³ ´
1 ∂ΠD
Since λ > 2 we have that ∂Eθj (m2j )
> 0. >From Lemma A.3-ii we have that Eθj m2j increases
with Nj and therefore ΠD increases as the number of elements Nj in the partition of Manager j
increases. ¥

Proof of Proposition 4: a.-Centralization: >From (28) and (29) and Lemma A2-iii we have
that
¡ ¡ ¢ ¡ ¢¢
ΠC = − AC σ 21 + σ 22 + (1 − AC ) SC σ 21 + σ 22 .

38
 b.-Decentralization: From (30) and (31) and Lemma A2-iii we have that
¡ ¡ ¢ ¡ ¢¢
ΠD = −ΠD = − AD σ 21 + σ 22 + BD SD σ 21 + σ 22 . ¥

Proof of Lemma 1: Substituting (3) and (4) into the definitions of ALC and CLC and making
use of Lemma A2 gives
µ ¶
1 + 8δ (1 + δ) ¡ ¢
ALC = 1 − 2 (1 − SC ) σ 21 + σ 22 (32)
(4δ + 1)

and
2δ ¡ 2 2
¢
CLC = (1 − S C ) σ 1 + σ 2 . (33)
(1 + 4δ)2
Similarly, substituting (6) and (7) into the definitions of ALD and CLD and making use of Lemma
A2 we obtain
à ¡ 2 ¢ !
1 2δ − λ2 ¡ 2 ¢
ALD = δ 2 2 − 2 2 (1 − SD ) σ 1 + σ 22 (34)
(λ + δ) (λ + δ) (λ + 2δ)

and µ ¶
2λ2 δ δ (2λ + 3δ) ¡ 2 ¢
CLD = 1− 2 (1 − SD ) σ 1 + σ 22 . (35)
(λ + δ)2 (λ + 2δ)
Subtracting (34) from (32) shows that ALC − ALD is equal to
¡ ¢¡ ¢
λδ (2λ − 1) ω 0 + δω1 + δ 2 ω 2 + δ 3 ω 3 + δ 4 ω4 σ 21 + σ 22
, (36)
(4δ + 1)2 (3λ + δ (8λ − 1)) (λ (5λ − 1) + δ (8λ − 1)) (λ + δ) (λ + 2δ)
¡ ¢ ¡ ¢
where ω0 ≡ λ2 (5λ − 1), ω 1 ≡ λ 33λ + 100λ2 − 7 , ω 2 ≡ 2 −9λ + 240λ2 + 100λ3 − 5 , ω 3 ≡
¡ ¢
2 174λ + 460λ2 − 43 and ω4 ≡ 16 (53λ − 10). It is straightforward to verify that ω i > 0, i =
1, ..., 4, for all λ ∈ [1/2, 1]. Inspection of (36) then shows that ALC − ALD = 0 if either δ = 0 or
λ = 1/2 and that ALC − ALD > 0 otherwise.
Next, subtracting (33) from (35) shows that CLD − CLC is equal to
¡ ¢¡ ¢
2λδ 2 (2λ − 1) γ 0 + δγ 1 + δ 22 γ + δ 33 γ σ 21 + σ 22
, (37)
(4δ + 1)2 (λ (5λ − 1) + δ (8λ − 1)) (3λ + δ (8λ − 1)) (λ + δ) (λ + 2δ)
¡ ¢ ¡ ¢
where γ 0 ≡ λ2 (65λ − 7), γ 1 ≡ λ 107λ + 280λ2 − 17 , γ 2 ≡ 2 −5λ + 224λ2 + 160λ3 − 3 and
γ 3 ≡ 4 (14λ + 3) (8λ − 1). It is straightforward to verify that γ i > 0, i = 1, 2, 3, for all λ ∈
[1/2, 1]. Inspection of (37) then shows that CLD − CLC = 0 if either δ = 0 or λ = 1/2 and that
CLD − CLC > 0 otherwise. ¥

39
Proof of Proposition 5: Using the expressions in Proposition 4 we have that ΠC − ΠD is given
by ¡ 2 ¢
σ 1 + σ 22 λδ (2λ − 1) f
,
(4δ + 1) (3λ + δ (8λ − 1)) (λ (5λ − 1) + δ (8λ − 1)) (λ + δ) (λ + 2δ)
where
¡ ¢
f ≡ 2δ 3 (4λ − 1) (28λ − 17) + 2δ 2 (5λ − 1) −3λ + 16λ2 − 5
¡ ¢
+λδ −51λ + 50λ2 + 7 − λ2 (5λ − 1) .

Note that the denominator is strictly positive for all δ ∈ [0, ∞) and λ ∈ [1/2, 1]. Thus, ΠC − ΠD
is continuous in δ ∈ [0, ∞) and λ ∈ [1/2, 1]. Also, ΠC − ΠD = 0 if either δ = 0, λ = 1/2 or f = 0.
Let λ (δ) be the values of λ ∈ [1/2, 1] which solve f = 0 for δ ∈ [0, ∞); λ (δ) is plotted in Figure
¡ ¢
5. Note that limδ→∞ λ (δ) = 17/28 and that λ δ = 1, where δ ' 0.192 57 is implicitly defined by
³ ´ £ ¢
2 3
3δ + 32δ + 33δ − 2 = 0. Note also that λ (δ) is decreasing in δ ∈ δ, ∞ .
Part (i.): Since λ (δ) is decreasing in δ it is sufficient to show that for λ = 1, ΠC − ΠD <
¡ ¢
0 if δ ∈ 0, δ . To see this note that sign (ΠC − ΠD ) = signf and that for λ = 1, f =
¡ ¢ ¡ ¢
2 3δ + 32δ 2 + 33δ 3 − 2 . It can be seen that f < 0 if δ ∈ 0, δ .
¡ ¢
Part (ii.): It is sufficient to show that for λ = 1, ΠC − ΠD > 0 if δ ∈ δ, ∞ . That this is
so follows immediately from the definition of δ and the facts that sign (ΠC − ΠD ) = signf and
¡ ¢
f = 2 3δ + 32δ 2 + 33δ 3 − 2 for λ = 1.
Part (iii.): Taking the limits of (14) and (15) gives:

5λ − 1 ¡ 2 ¢
lim ΠC = lim ΠD = − σ 1 + σ 22 . ¥
δ→∞ δ→∞ 8λ − 1

Proof of Proposition 6: Using the definitions of SC and SD we have that SC = SD /2 if and

only if δ = 0 or ¡ ¢
(2λ − 1) −3λ2 − δ 2 − 4λδ + 8λδ 2 + 2λ2 δ
¡ ¢ = 0.
2 (3λ − δ + 8λδ) −λ − δ + 5λ2 + 8λδ
Note that the terms in the denominator are not equal to zero for any λ ∈ [1/2, 1] and δ ∈ [0, ∞).
¡ ¢
Thus, SC = SD /2 if either λ = 1/2 or −3λ2 − δ 2 − 4λδ + 8λδ 2 + 2λ2 δ = 0. One solution to this
quadratic equation is given by δ 1 as defined in the Proposition and the other solution is negative.
Finally, it is straightforward to verify that for λ = 1, SC < SD /2 if δ ∈ (0, δ 1 ). Thus, SC ≤ SD /2
for all δ ∈ [0, δ 1 ] and SC ≥ SD /2 for all δ ∈ [δ 1 , ∞). ¥

40
10 Appendix B - Sequential Decision Making

Decision Making: In the last stage of the game Manager 2 chooses d2 to maximize his expected
utility E2 [ (1 − λ)π 1 + λπ 2 | d1 ]. The optimal decision that solves this problem is given by

λ δ
dS2 ≡ θ2 + dS . (38)
λ+δ λ+δ 1
At the previous stage Manager 1 chooses d1 to maximize his expected profits E1 [ λπ 1 + (1 − λ)π 2 | m].
The optimal decision is given by

λ (λ + δ)2 λ2 + δ (1 − λ)
dS1 ≡ θ 1 + δ E1 [θ2 | m] . (39)
λ3 + 3λ2 δ + δ 2 λ3 + 3λ2 δ + δ 2
Communication: Let μ2 (m2 | θ2 ) be the probability with which Manager 2 sends message m2 ,
let d1 (m2 ) and d2 (m2 ) be the decision rules that map messages into decisions and let g1 (θ2 | m2 )
be the belief function which gives the probability of θ2 conditional on observing m2 . We can now
state the following proposition which characterizes the finite communication equilibria when δ > 0.

PROPOSITION A1 (Communication Equilibria). If δ ∈ (0, ∞), then for every positive integer N2
there exists at least one equilibrium ( μ2 (·) , d1 (·), d2 (·), g1 (·)), where

i. μ2 (m2 | θ2 ) is uniform, supported on [a2,i−1 , a2,i ] if θ2 ∈ (a2,i−1 , a2,i ),

ii. g1 (θ2 | m2 ) is uniform supported on [a2,i−1 , a2,i ] if m2 ∈ (a2,i−1 , a2,i ),

iii. a2,i+1 − a2,i = a2,i − a2,i−1 + 4bS a2,i for i = 1, ..., N2 − 1,

a2,−(i+1) − a2,−i = a2,−i − a2,−(i−1) + 4bS a2,−i for i = 1, ..., Nj − 1,

¡ ¡ ¢¢ ¡ ¡ ¢¢
where bS ≡ (2λ − 1) (λ + δ) λ2 + δ / (λ (1 − λ) + δ) λ2 + δ (1 − λ) and
iv. dj (m) = dSj , j = 1, 2, where dSj is given by (38) and (39).
Moreover, all other finite equilibria have relationships between θ1 and θ2 and the managers’ choices
of d1 and d2 that are the same as those in this class for some value of N2 ; they are therefore
economically equivalent.

Proof: The proof is analogous to the proof of Proposition 1. Details are available from the authors
upon request. ¥

PROPOSITION A2 (Efficiency). The limit of strategy profiles and beliefs (μ2 (·) , d1 (·), d2 (·), g1 (·))
as N2 → ∞ is a Perfect Bayesian Equilibrium of the communication game. In this equilibrium
the total expected profits E [π 1 + π 2 ] are higher than in any other equilibrium.

41
Proof: The proof is analogous to the proof of Proposition 2. Details are available from the authors
upon request. ¥

In the remaining analysis we focus on the efficient equilibrium.

LEMMA A1. In the most efficient equilibrium in which N2 → ∞ the residual variance is given by
h i
E (θ2 − E [θ2 |m2 ])2 = SS σ 22 ,

where SS = bS /(3 + 4bS ).

Proof: The proof is analogous to the proof of Lemma 1. Details are available from the authors
upon request. ¥

PROPOSITION A4 (Organizational Performance). Under Decentralization with sequential decision

making the expected profits are given by
¡ ¡ ¢ ¢
ΠS = − (AD + X) σ 21 + σ 22 + (BD − X) SS σ 22 , (40)

where AD and BD are defined in (16) and

2λ4 + λ2 (6λ + 1) δ + 2λ (2 + λ) δ 2 + 2δ 3
X ≡ δ 3 (2λ − 1)2 ¡ ¢2 .
(λ + 2δ)2 λ3 + δ 2 + 3λ2 δ

Proof: The proof is analogous to the proof of Proposition 4. Details are available from the authors
upon request. ¥

We can now prove Proposition 7 which is stated in the main text.

Proof of Proposition 7: i. Applying l’Hopital’s Rule to (14) and (40) and using the assumption
that σ 21 = σ 22 = σ 2 we obtain

8λ (4λ − 1) (2λ − 1)2 2

lim ΠC − lim ΠS = σ
δ→∞ δ→∞ (8λ − 1) (5λ − 1)
which is strictly positive for any λ > 1/2.
ii. Taking the derivative of (14) and (40) and using the assumption that σ 21 = σ 22 = σ 2 we
obtain
d (ΠS − ΠC ) 8δ
= σ2
dλ 3 (1 + 2δ) (1 + 4δ)
which is strictly positive for all finite δ > 0. ¥

42
11 Appendix C - Different Needs for Coordination

Since allowing for differences in the needs for coordination only requires adding a parameter in the
main model, we do not replicate the full analysis here. Instead we merely state the key expressions
and use them to prove Proposition 8. The derivation of these expressions and their interpretation
are exactly as in the main model.

11.1 Centralization

The decisions are now given by

µ ¶
C 1
d1 ≡ ((1 + δ 1 + δ 2 ) EH [θ1 | m] + (δ 1 + δ 2 ) EH [θ2 | m])
1 + 2 (δ 1 + δ 2 )
µ ¶
C 1
d2 ≡ ((δ 1 + δ 2 ) EH [θ1 | m] + n (1 + δ 1 + δ 2 ) EH [θ2 | m]) .
1 + 2 (δ 1 + δ 2 )

The residual variance of θ1 is given by SC,1 σ 21 and that of θ2 is given by SC,2 σ 22 , where SC,j ≡
bC,j / (3 + 4bC,j ), j = 1, 2, and
³ ´
(2λ − 1) δ 2 + (δ 1 + δ 2 )2
bC,1 =
δ 2 + (δ 1 + δ 2 )2 + λ (1 + 3δ 1 + δ 2 )
³ ´
(2λ − 1) δ 1 + (δ 1 + δ 2 )2
bC,2 = .
δ 1 + (δ 1 + δ 2 )2 + λ (1 + δ 1 + 3δ 2 )

The expected profits are given by

µ ¶
2 δ1 + δ2 1 + δ1 + δ2
ΠC = −σ 2 + (S1 + S2 ) . (41)
1 + 2 (δ 1 + δ 2 ) 1 + 2 (δ 1 + δ 2 )
Applying l’Hopital’s Rule gives
5λ − 1 2
lim ΠC = −2 σ . (42)
δ 1 →∞ 8λ − 1
We can also use (41) to evaluate dΠC /dλ for λ = 1/2:
dΠC 4 (δ 1 + δ 2 )
=− σ2 . (43)
dλ 3 1 + 2 (δ 1 + δ 2 )

43
11.2 Decentralization

The decisions under Decentralization are now given by

λθ1 ((1 − λ) δ 1 + λδ 2 ) (λδ 1 + (1 − λ) δ 2 )

dD
1 = + E2 [θ1 | m]
λ + λδ 1 + (1 − λ) δ 2 (λ + δ 1 + δ 2 ) (λ + λδ 1 + (1 − λ) δ 2 )
λδ 1 + (1 − λ) δ 2
+ E1 [θ2 | m]
λ + δ1 + δ2
λθ2 (1 − λ) δ 1 + λδ 2
dD
2 = + E2 [θ1 | m]
λ + (1 − λ) δ 1 + λδ 2 λ + δ1 + δ2
((1 − λ) δ 1 + λδ 2 ) (λδ 1 + (1 − λ) δ 2 )
+ E1 [θ2 | m] .
(λ + δ 1 + δ 2 ) (λ + (1 − λ) δ 1 + λδ 2 )

The residual variance of θ1 is given by SD,1 σ 21 and that of θ2 is given by SD,2 σ 22 , where SD,j ≡
bC,j / (3 + 4bC,j ), j = 1, 2, and

(2λ − 1) δ 1 (λ + λδ 1 + (1 − λ) δ 2 )
b1 =
(λ (1 − λ) + λδ 1 + (1 − λ) δ 2 ) ((1 − λ) δ 1 + λδ 2 )
(2λ − 1) δ 2 (λ + λδ 2 + (1 − λ) δ 1 )
b2 = .
(λ (1 − λ) + λδ 2 + (1 − λ) δ 1 ) ((1 − λ) δ 2 + λδ 1 )
The expected profits are given by
h¡ ¢2 ¡ D ¢2 ¡ ¢ i
D 2
ΠD = −E dD
1 − θ1 + d2 − θ2 + (δ 1 + δ 2 ) dD
1 − d2 , (44)

where
h¡ ¢2 i
EdD
1 − θ 1
Ã
(δ 2 + λδ 1 − λδ 2 )2 (λδ 1 + (1 − λ) δ 2 )2
= σ2 2 − S2
(λ + δ 1 + δ 2 )2 (λ + δ 1 + δ 2 )2
!
((1 − λ) δ 1 + λδ 2 ) (λδ 1 + (1 − λ) δ 2 )2 (2λ + (1 + λ) δ 1 + (2 − λ) δ 2 )
+ S1
(λ + δ 1 + δ 2 )2 (λ + λδ 1 + (1 − λ) δ 2 )2

h i
E (d2 − θ2 )2
Ã
(−δ 1 + λδ 1 − λδ 2 )2 ((1 − λ) δ 1 + λδ 2 )2
= σ2 2 − S1
(λ + δ 1 + δ 2 )2 (λ + δ 1 + δ 2 )2
!
((1 − λ) δ 1 + λδ 2 )2 (λδ 1 + (1 − λ) δ 2 ) (2λ + (2 − λ) δ 1 + (1 + λ) δ 2 )
+ S2
(λ + δ 1 + δ 2 )2 (λ + (1 − λ) δ 1 + λδ 2 )2

44
 h¡ ¢ i
D 2
E dD
1 − d2
µ
2 λ2 λ2 (2λ + (1 + λ) δ 1 + (2 − λ) δ 2 ) ((1 − λ) δ 1 + λδ 2 )
= σ 2 + S1
(λ + δ 1 + δ 2 )2 (λ + δ 1 + δ 2 )2 (λ + λδ 1 + (1 − λ) δ 2 )2
¶
λ2 (2λ + (2 − λ) δ 1 + (1 + λ) δ 2 ) (λδ 1 + (1 − λ) δ 2 )
+ S2 .
(λ + δ 1 + δ 2 )2 (λ + (1 − λ) δ 1 + λδ 2 )2

Applying l’Hopital’s Rule gives

8λ3 − 9λ2 + 6λ − 1 2
lim ΠD = −2 σ . (45)
δ 1 →∞ 5λ − 1
We can also use (44) to evaluate dΠC /dλ for λ = 1/2:

dΠD 8 (δ 1 + δ 2 )2
=− σ2 . (46)
dλ 3 (1 + 2 (δ 1 + δ 2 ))2

We can now prove Proposition 8 which is stated in the main text.

Proof of Proposition 8: i. Using (42) and (45) we obtain

(2λ − 1)2
lim ΠC − lim ΠS = 8λ (4λ − 1) σ2
δ→∞ δ→∞ (8λ − 1) (5λ − 1)
which is strictly positive for any λ > 1/2.
ii. Using (43) and (46) we find that the difference in the derivatives at λ = 1/2 is given by
d (ΠC − ΠS ) 4 (δ 1 + δ 2 )
= σ 2 for λ = 1/2
dλ 3 (1 + 2 (δ 1 + δ 2 ))2

which is strictly positive for all finite δ > 0. ¥

45
12 Appendix D - Different Division Sizes

Since allowing for different division sizes only requires adding a parameter in the main model, we
do not replicate the full analysis here. Instead we merely state the key expressions and use them
to prove Proposition 9. The derivation of these expressions and their interpretation are exactly as
in the main model.

12.1 Centralization

Let β ≡ (1 − α). Then decisions are given by

µ ¶
C 1
d1 ≡ (α (β + δ) EH (θ1 | m) + βδEH (θ2 | m))
αβ + δ
µ ¶
C 1
d2 ≡ (αδEH (θ1 | m) + β (α + δ) EH (θ2 | m)) .
αβ + δ
The residual variance of θ1 is given by SC,1 σ 21 and that of θ2 is given by SC,2 σ 22 , where SC,j ≡
bC,j / (3 + 4bC,j ), j = 1, 2, and
¡ ¢
βδ (2λ − 1) β 2 + δ
bC,1 = ¡ ¢ ¡ ¢
αλ β 2 + δ 2 + β δ + β 2 (1 − λ) δ + αβλ (2 + β) δ
¡ ¢
αδ (2λ − 1) α2 + δ
bC,2 = ¡ ¢ .
α (δ + α2 ) (1 − λ) δ + βλ α2 + δ 2 + αβλ (2 + α) δ
The expected profits are given by
h¡ ¢2 ¡ C ¢2 ¡ C ¢ i
C 2
ΠC = −E dC 1 − θ 1 + d2 − θ 2 + 2δ d1 − d2 ,

where
h i µ ¶
2δ 2 β 2 αβ + (2 − α) δ δ2β2
E (d1 − θ1 )2 = σ 2 + α (β + δ) S1 − S2
(αβ + δ)2 (αβ + δ)2 (αβ + δ)2
h i µ ¶
2 2 2α2 δ 2 α2 δ 2 αβ + (1 + α) δ
E (d2 − θ2 ) = σ − S1 + β (α + δ) S2
(αβ + δ)2 (αβ + δ)2 (αβ + δ)2
h i µ ¶
2 2 2α2 β 2 α2 β 2
E (d1 − d2 ) = σ − (S1 + S2 ) .
(αβ + δ)2 (αβ + δ)2
Applying l’Hopital’s Rule we find that
−2α (1 − α) (8λ − 1) (5λ − 1)
lim ΠC = σ2 . (47)
δ→∞ (5λ − 1 − α (2λ − 1)) (3λ + (2λ − 1) α)
Also, differentiating we find that at λ = 1/2
dΠC 8 α (1 − α) δ 2
=− σ . (48)
dλ 3 α (1 − α) + δ

46
12.2 Decentralization

The decisions are now given by

¡ ¡ ¢¢
αλθ1 + δ (αλ + β (1 − λ)) E1 dD
2 |m
dD
1 =
αλ (1 + δ) + βδ (1 − λ)
¡ ¡ ¢¢
βλθ2 + δ (α (1 − λ) + βλ) E2 dD
1 |m
dD
2 =
βλ (1 + δ) + αδ (1 − λ)
where β ≡ (1 − α) and
£ ¤ (α (αδ (1 − λ) + βλ (1 + δ)) E2 [θ1 | m] + βδ (αλ + β (1 − λ)) E1 [θ2 | m])
E2 dD1 |m = ¡ ¢
α2 + β 2 δ (1 − λ) + αβλ (1 + 2δ)
£ ¤ (αδ (α (1 − λ) + βλ) E2 [θ1 | m] + β (αλ (1 + δ) + βδ (1 − λ)) E1 [θ2 | m])
E1 dD2 |m = ¡ ¢ .
α2 + β 2 δ (1 − λ) + αβλ (1 + 2δ)
The residual variance of θ1 is given by SD,1 σ 21 and that of θ2 is given by SD,2 σ 22 , where SD,j ≡
bD,j / (3 + 4bC,j ), j = 1, 2, and
αβ (2λ − 1) (αλ (1 + δ) + β (1 − λ) δ)
bD, 1 = ³ ´
(α (1 − λ) + βλ) β 2 (1 − λ)2 δ + α2 λ2 δ + αβλ (1 + 2δ) (1 − λ)
αβ (2λ − 1) (α (1 − λ) δ + βλ (1 + δ))
bD,2 = ³ ´
(β + αλ − βλ) α2 (1 − λ)2 δ + β 2 λ2 δ + αβλ (2δ + 1) (1 − λ)

The expected profits are given by

h¡ ¢2 ¡ D ¢2 ¡ ¢ i
D 2
ΠD = −E dD 1 − θ1 + d2 − θ2 + 2δ dD
1 − d2 ,

where
h¡ ¢2 i
E dD
1 − θ 1
Ã
σ2 2δ 2 β 2 (αλ (1 + δ) + β (1 − λ) δ)2 (αλ + β (1 − λ))2
= ¡¡ ¢ ¢2
α2 + β 2
(1 − λ) δ + αβλ (1 + 2δ) (αλ (1 + δ) + β (1 − λ) δ)2
¡¡ ¢ ¢
αδ 3 (α (1 − λ) + βλ) (αλ + β (1 − λ))2 α2 + 2β 2 (1 − λ) δ + αβλ (2 + 3δ) S1
+
(αλ (1 + δ) + β (1 − λ) δ)2
´
−β 2 δ 2 (αλ + β (1 − λ))2 S2
h¡ ¢2 i
E dD
2 − θ 2

σ2 ³
2 2 2 2 2 2
= ¡¡ ¢ ¢2 2δ α (α (1 − λ) + βλ) − α δ (α (1 − λ) + βλ) S1
2 2
α + β (1 − λ) δ + αβλ (1 + 2δ)
¡ ¡ ¢ ¢ !
βδ 3 (α (1 − λ) + βλ)2 (αλ + β (1 − λ)) δ 2α2 + β 2 (1 − λ) + αβλ (2 + 3δ) S2
+
(α (1 − λ) δ + βλ (1 + δ))2

47
 h¡ ¢2 i
E dD D
1 − d2

σ 2 λ2 ¡ 2 2
= ¡¡ ¢ ¢2 2α β
α2 + β 2 (1 − λ) δ + αβλ (1 + 2δ)
¡¡ ¢ ¢
α3 δ α2 + 2β 2 (1 − λ) δ + αβλ (3δ + 2) (α (1 − λ) + βλ) S1
+
(αλ (1 + δ) + β (1 − λ) δ)2
¡ ¡ ¢ ¢ !
β 3 δ (αλ (1 + δ) + βδ (1 − λ)) δ 2α2 + β 2 (1 − λ) + αβλ (2 + 3δ) (αλ + β (1 − λ)) S2
+ .
(αλ (1 + δ) + βδ (1 − λ)) (α (1 − λ) δ − βλ (1 + δ))2

Applying l’Hopital’s Rule we find that

³ ´
−2α (1 − α) 2α (1 − α) + 2λ2 (2α − 1)2 (3λ − 7) − 26αλ (1 − α) + 9λ − 1
lim ΠD = ³ ´ σ2. (49)
δ→∞ 2
1 − 2α (1 − α) − λ (2α − 1) (3λ (1 − λ) + α (1 − α) (6λ + 1) (2λ − 1))

Also, differentiating we find that at λ = 1/2

dΠD 16 1 − 2α (1 − α) 2
= − α (1 − α) δ 2 σ . (50)
dλ 3 (α (1 − α) + δ)2

We can now prove Proposition 9 which is stated in the main text.

Proof of Proposition 9: i. Using (47) and (49) gives

lim ΠC − lim ΠD
δ→∞ δ→∞
6αλ (1 − α) (2α − 1)2 (2λ − 1)
= ³ ´
1 − 2α (1 − α) − λ (2α − 1)2 (5λ − 1 − α (2λ − 1))
¡ ¡ ¢ ¢
α (1 − α) (2λ − 1) 42λ2 − 11λ + 1 + λ (1 − λ) (5λ − 1) 2
× σ
((2λ − 1) α + 3λ) (α (1 − α) (6λ + 1) (2λ − 1) + 3λ (1 − λ))
which is strictly positive if λ > 1/2 and α > 1/2.
ii. Using (48) and (50) we find that the difference in the derivatives at λ = 1/2 is given by

d (ΠD − ΠC ) 8 α (1 − α) (1 + 4δ) − δ 2
= α (1 − α) δ σ for λ = 1/2
dλ 3 (α (1 − α) + δ)2

which is strictly positive if à !

r
1 1
α< 1+ . ¥
2 1 + 4δ

48
References

[1] Aoki, M. 1986. Horizontal vs. Vertical Information Structure of the Firm. American Economic
Review, 76: 971-983.

[2] Alonso, R. and N. Matouschek. 2004. Relational Delegation. Mimeo, Northwestern.

[3] Athey, S. and J. Roberts 2001. Organizational Design: Decision Rights and Incentive
Contracts. American Economic Review Papers and Proceedings, 91: 200-205

[4] Aumann, R. and S. Hart 2003. Long Cheap Talk. Econometrica, 71, 1619-1660.

[5] Baliga, S. and S. Morris. 2002. Coordination, Spillovers, and Cheap Talk. Journal of
Economic Theory, 105: 450-68.

[6] Bartlett, C.A. 1989. Procter & Gamble Europe: Vizir Launch. Harvard Business School
Case 9-384-139.

[7] Battaglini, M. 2002. Multiple Referrals and Multidimensional Cheap Talk. Econometrica,
70(4): 1379-1401.

[8] Bolton, P. and M. Dewatripont 1994. The Firm as a Communication Network. Quarterly
Journal of Economics, CIX, 809-839.

[9] Bolton, P and J. Farrell. 1990. Decentralization, Duplication and Delay. Journal of
Political Economy, 98, 803-26.

[10] Chandler, A. 1977. The Visible Hand: The Managerial Revolution in American Business.
Cambridge: Belknap Press.

[11] Cremer, J., L. Garicano and A. Prat 2006. Language and the Theory of the Firm.
Quarterly Journal of Economics, forthcoming.

[12] Crawford, V. and J. Sobel. 1982. Strategic Information Transmission. Econometrica,

50:1431-145.

[13] Dessein, W. 2002. Authority and Communication in Organizations. Review of Economic

Studies, 69: 811—838.

[14] Dessein, W., L. Garicano and R. Gertner 2005. Organizing for Synergies. Mimeo,
Chicago GSB.

49
[15] Dessein, W. and T. Santos 2006. Adaptive Organizations. Journal of Political Economy,
forthcoming.

[16] Eccles, R.G. and P. Holland 1989. Jacobs Suchard: Reorganizing for 1992. Harvard
Business School Case 489106.

[17] Farrell, J. 1987. Cheap Talk, Coordination, and Entry. The RAND Journal of Economics.
18(1): 34-9

[18] Farrell, J. and R. Gibbons 1989. Cheap Talk Can Matter in Bargaining. Journal of Eco-
nomic Theory, 48(1): 221-37.

[19] Farrell, J. and M. Rabin 1996. Cheap Talk. The Journal of Economic Perspectives,
10(3):103-118.

[20] Friebel, G. and M. Raith 2006. Resource Allocation and Firm Scope. Mimeo, University
of Toulouse.

[21] Garicano, L. 2000. Hierarchies and the Organization of Knowledge in Production. Journal
of Political Economy, October, 874-904.

[22] Gertner, R. 1999. Coordination, Dispute Resolution, and the Scope of the Firm, Mimeo,
GSB Chicago.

[23] Grossman, S. and O. Hart 1986. The Costs and Benefits of Ownership: A Theory of
Vertical and Lateral Integration,” Journal of Political Economy, 94, 691-719.

[24] Hannan, M., J. Podolny and J. Roberts 1999. The Daimler Chrysler Commercial Vehicles
Division. Stanford Business School Case IB-27.

[25] Hart, O. and B. Holmstrom 2002. A Theory of Firm Scope. Mimeo, MIT.

[26] Hart, O. and J. Moore 1990. Property Rights and the Nature of the Firm. Journal of
Political Economy, 98, 1119-58.

[27] Harris, M. and A. Raviv 2005. Allocation of Decision-Making Authority. Review of Finance,
September 2005, 9, 353-83.

[28] Hart, O. and J. Moore 1990. Property Rights and the Nature of the Firm. Journal of
Political Economy, 98: 1119-58.

50
[29] Hart, O. and J. Moore 2005. On the Design of Hierarchies: Coordination versus Special-
ization. Journal of Political Economy, 113(4): 675-702.

[30] Hayek, F. 1945. The Use of Knowledge in Society. American Economic Review, 35(4): 519-
530.

[31] Jensen, m. and W. MecKling 1992. Specific and General Knowledge, and Organizational
Structure. In Contract Economics, (Lars Werin and Hans Wijkander, eds.), Oxford: Blackwell.

[32] Kawamura, K. 2006. Weighted Cheap Talk: Anonymity, Garbling, and Overconfidence in
Communication. Mimeo, University of Oxford.

[33] Krishna, V. and J. Morgan 2001. A Model of Expertise. Quarterly Journal of Economics,
116: 747-775.

[34] Krishna, V. and J. Morgan 2004. The Art of Conversation: Eliciting Information from
Experts Through Multi-Stage Communication. Journal of Economic Theory, 117: 147-179

[35] Lockwood, B. 2005. Fiscal Decentralization: A Political Economy Perspective, in The Hand-
book of Fiscal Federalism (E. Ahmad and G. Brosio eds.), Edward Elgar, forthcoming.

[36] Marshak J. and R. Radner 1972. Economic Theory of Teams. Yale University Press, New
Haven and London.

[37] Marino, A. and J. Matsusaka 2005. Decision Processes, Agency Problems, and Informa-
tion: An Economic Analysis of Capital Budgeting Procedures. Review of Financial Studies,
18(1): 301-325.

[38] Melumad, N. and T. Shibano 1991. Communication in Settings with No Transfers. Rand
Journal of Economics, 22: 173-198.

[39] Milgrom, P. and J. Roberts. 1992. Economics, Organization, and Management. Prentice
Hall, Englewood Cliffs, New Jersey.

[40] Montgomery, C.A. and D. Magnani. 2001. PepsiCo’s Restaurants. Harvard Business
School Case 9-794-078.

[41] Mookherjee, D. 2006. Decentralization, Hierarchies, and Incentives: A Mechanism Design

Perspective. Journal of Economic Literature, XLIV, 367-390.

51
[42] Oates, W. 1999. An Essay on Fiscal Federalism. Journal of Economic Literature, 37: 1120-
1149.

[43] Ozbas, O. 2005. Integration, Organizational Processes, and Allocation of Resources. Journal
of Financial Economics, 75(1), 201-242.

[44] Pfeffer, J. 2004. Human Resources at the AES Corp.: The Case of the Missing Department,
Stanford Business School HR3.

[45] Radner, R. 1993. The Organization of Decentralized Information Processing. Econometrica,

61: 1109—1146.

[46] Rantakari, H. 2006. Coordination and Strategic Communication. Mimeo, MIT.

[47] Qian, Y., G. Roland and C. Xu. 2006. Coordination and Experimentation in M-Form and
U-Form Organizations. Journal of Political Economy, 114: 366—402.

[48] Sloan, A.P., Jr. 1964. My Years With General Motors. New York: Doubleday.

[49] Stein, J. 1997. Internal Capital Markets and the Competition for Corporate Resources. Jour-
nal of Finance, 52: 111-133.

[50] Stein, J. 2002. Information Production and Capital Allocation: Decentralized versus Hierar-
chical Firms, Journal of Finance, 57: 1891-1921.

[51] Van Zandt, T. 1999. Decentralized Information Processing in the Theory of Organizations.
In Contemporary Economic Issues, Vol. 4: Economic Design and Behavior (Murat Sertel ed.),
London: MacMillan.

[52] Wulf, J. 2005. Influence and Inefficiency in the Internal Capital Market. Mimeo, University
of Pennsylvania.

52